TAO: A Large-Scale Benchmark for Tracking Any Object

05/20/2020 ∙ by Achal Dave, et al. ∙ 18

For many years, multi-object tracking benchmarks have focused on a handful of categories. Motivated primarily by surveillance and self-driving applications, these datasets provide tracks for people, vehicles, and animals, ignoring the vast majority of objects in the world. By contrast, in the related field of object detection, the introduction of large-scale, diverse datasets (e.g., COCO) have fostered significant progress in developing highly robust solutions. To bridge this gap, we introduce a similarly diverse dataset for Tracking Any Object (TAO). It consists of 2,907 high resolution videos, captured in diverse environments, which are half a minute long on average. Importantly, we adopt a bottom-up approach for discovering a large vocabulary of 833 categories, an order of magnitude more than prior tracking benchmarks. To this end, we ask annotators to label objects that move at any point in the video, and give names to them post factum. Our vocabulary is both significantly larger and qualitatively different from existing tracking datasets. To ensure scalability of annotation, we employ a federated approach that focuses manual effort on labeling tracks for those relevant objects in a video (e.g., those that move). We perform an extensive evaluation of state-of-the-art trackers and make a number of important discoveries regarding large-vocabulary tracking in an open-world. In particular, we show that existing single- and multi-object trackers struggle when applied to this scenario in the wild, and that detection-based, multi-object trackers are in fact competitive with user-initialized ones. We hope that our dataset and analysis will boost further progress in the tracking community.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 4

page 9

page 13

page 19

page 25

page 28

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

A key component in the success of modern object detection methods was the introduction of large-scale, diverse benchmarks, such as MS COCO [40] and LVIS [28]. By contrast, multi-object tracking datasets tend to be small [42, 60], biased towards short videos [70], and, most importantly, focused on a very small vocabulary of categories [42, 60, 64] (see Table 1). As can be seen from Figure 1, they predominantly target people and vehicles. Due to the lack of proper benchmarks, the community has shifted towards solutions tailored to the few videos used for evaluation. Indeed, Bergmann et al. [5] have recently and convincingly demonstrated that simple baselines perform on par with state-of-the-art (SOTA) multi-object trackers.

Figure 1: (left) Super-category distribution in existing multi-object tracking datasets compared to TAO and COCO [40]. Previous work focused on people, vehicles and animals. By contrast, our bottom-up category discovery results in a more diverse distribution, covering many small, hand-held objects that are especially challenging from the tracking perspective. (right) Wordcloud of TAO categories, weighted by number of instances, and colored according to their supercategory.

In this work we introduce a large-scale benchmark for Tracking Any Object (TAO). Our dataset features 2,907 high resolution videos captured in diverse environments, which are 30 seconds long on average, and has tracks labeled for 833 object categories. We compare the statistics of TAO to existing multi-object tracking benchmarks in Table 1 and Figure 1, and demonstrate that it improves upon them both in terms of complexity and in terms of diversity (see Figure 2 for representative frames from TAO). Collecting such a dataset presents three main challenges: (1) how to select a large number of diverse, long, high-quality videos; (2) how to define a set of categories covering all the objects that might be of interest for tracking; and (3) how to label tracks for these categories at a realistic cost. Below we summarize our approach for addressing these challenges. A detailed description of dataset collection is provided in Section 4. Existing datasets tend to focus on one or just a few domains when selecting the videos, such as outdoor scenes in MOT [42], or road scenes in KITTI [25]. This results in methods that fail when applied in the wild. To avoid this bias, we construct TAO with videos from as many environments as possible. We include indoor videos from Charades [56], movie scenes from AVA [27], outdoor videos from LaSOT [21], road-scenes from ArgoVerse [14], and a diverse sample of videos from HACS [73] and YFCC100M [58]. We ensure all videos are of high quality, with the smallest dimension larger or equal to 480px, and contain at least 2 moving objects. creftypecap 1 reports the full statistics of the collected videos, showing that TAO provides an evaluation suite that is significantly larger, longer, and more diverse than prior work. Note that TAO contains fewer training videos than recent tracking datasets, as we intentionally dedicate the majority of videos for in-the-wild benchmark evaluation, the focus of our effort.

Dataset Classes
Videos
Eval. Train
Avg
length (s)
Tracks
/ video
Min
resolution
Ann.
fps
Total Eval
length (s)
MOT17 [42] 1 7 7 35.4 112 640x480 30 248
KITTI [25] 2 29 21 12.6 52 1242x375 10 365
UA-DETRAC [64] 4 40 60 56 57.6 960x540 5 2,240
ImageNet-Vid  [52] 30 1,314 4,000 10.6 2.4 480x270 25 13,928
YTVIS  [70] 40 645 2,238 4.6 1.7 320x240 5 2,967
TAO (Ours) 833 2,407 500 36.8 5.9 640x480 1 88,605
Table 1: Statistics of major multi-object tracking datasets. TAO is by far the largest dataset in terms of the number of categories, and the total duration of videos used for evaluation. In addition, we ensure that each video is challenging (long, containing several moving objects) and of high quality.

Given the selected videos, we must choose what to annotate. Most datasets are constructed with a top-down approach, where categories of interest are pre-defined by benchmark curators. That is, curators first select the subset of categories deemed relevant for the task, and then collect images or videos expressly for these categories [19, 40, 59]. This approach naturally introduces curator bias. An alternative strategy is bottom-up, open-world discovery of what objects are present in the data. Here, the vocabulary emerges post factum [27, 28, 74], an approach that dates back to LabelMe [53]. Inspired by this line of work, we devise the following strategy to discover an ontology of objects relevant for tracking: first annotators are asked to label all objects that either move by themselves or are moved by people. They then give names to the labeled objects, resulting in a vocabulary that is not only significantly larger, but is also qualitatively different from that of any existing tracking dataset (see Figure 1). To facilitate training of object detectors, that can be later used by multi-object trackers on our dataset, we encourage annotators to choose categories that exists in the LVIS dataset [28]. If no appropriate category can be found in the LVIS vocabulary, annotators can provide free-form names (see Section 4.2 for details). Exhaustively labeling tracks for such a large collection of objects in 2,907 long videos is prohibitively expensive. Instead, we extend the federated annotation approach proposed in [28] to the tracking domain. In particular, we ask the annotators to label tracks for up to 10 objects in every video. We then separately collect exhaustive labels for every category for a subset of videos, indicating whether all the instances of the category have been labeled in the video. During evaluation of a particular category, we use only videos with exhaustive labels for computing precision and all videos for computing recall. This allows us to reliably measure methods’ performance at a fraction of the cost of exhaustively annotating the videos. We use the LVIS federated mAP metric [28] for evaluation, replacing 2D IoU with 3D IoU [70]. For detailed comparisons, we further report the standard MOT challenge [42] metrics in creftypecap 0.B.2.

Figure 2: Representative frames from TAO, showing videos sourced from multiple domains with annotations at two different timesteps.

Equipped with TAO, we set out to answer several questions about the state of the tracking community. In particular, in Section 5 we report the following discoveries: (1) SOTA trackers struggle to generalize to a large vocabulary of objects, particularly for infrequent object categories in the tail; (2) while trackers work significantly better for the most-explored category of people, tracking people in diverse scenarios (e.g., frequent occlusions or camera motion) remains challenging; (3) when scaled to a large object vocabulary, multi-object trackers become competitive with user-initialized trackers, despite the latter being provided with a ground truth initializations. We hope that these insights will help to define the most promising directions for future research.

2 Related work

The domain of object tracking is subdivided based on the way the tracks are initialized. Our work falls into the multi-object tracking category, where all the objects out of a fixed vocabulary of classes have to be detected and tracked. Other formulations include user-initialized tracking, and saliency-based tracking. In the remainder of this section we will first review the most relevant benchmarks datasets in each of these areas, and then discuss SOTA methods for multi-object and user-initialized tracking.

2.1 Benchmarks

Multi-object tracking (MOT) is the task of tracking an unknown number of objects from a known set of categories. Most MOT benchmarks [23, 25, 42, 64] focus on either people or vehicles (see Figure 1), motivated by surveillance and self-driving applications. Moreover, they tend to include only a few dozen videos, captured in outdoor or road environments, encouraging methods that are overly adapted to the benchmark and do not generalize to different scenarios (see Table 1). In contrast, TAO focuses on diversity both in the category and visual domain distribution, resulting in a realistic benchmark for tracking any object. Several works have attempted to extend the MOT task to a wider vocabulary of categories. In particular, the ImageNet-Vid [52] benchmark provides exhaustive trajectories annotations for objects of 30 categories in 1314 videos. While this dataset is both larger and more diverse that standard MOT benchmarks, videos tend to be relatively short and the categories cover only animals and vehicles. The recent YTVIS dataset [70] has the most broad vocabulary to date, covering 40 classes, but the majority of the categories still correspond to people, vehicles and animals. Moreover, the videos are 5 seconds long on average, making the tracking problem considerably easier in many cases. Unlike previous work, we take a bottom-up approach for defining the vocabulary. This results in not only the largest set of categories among MOT datasets to date, but also in a qualitatively different category distribution. In addition, our dataset is over 7 times larger than YTVIS in the number of frames. The recent VidOR dataset [55] explores Video Object Relations, including tracks for a large vocabulary of objects. But, since ViDOR focuses on relations rather than tracks, object trajectories tend to be missing or incomplete, making it hard to repurpose for tracker benchmarking. In contrast, we ensure TAO maintains high quality for both accuracy and completeness of labels (see Appendix 0.A.1 for a quantitative analysis). Finally, several recent works have proposed to label masks instead of bounding boxes for benchmarking multi-object tracking [60, 70]. In collecting TAO we made a conscious choice to prioritize scale and diversity of the benchmark over pixel-accurate labeling. Instance mask annotations are significantly more expensive to collect than bounding boxes, and we show empirically that tracking at the box level is already a challenging task that current methods fail to solve. User-initialized tracking forgoes a fixed vocabulary of categories altogether and instead relies on the user to provide bounding box annotations for the objects that need to be tracked at test time [21, 32, 36, 59, 67]. The benchmarks in this category tend to be larger and more diverse than their MOT counterparts, but most of them still offer a tradeoff between the number of videos in the benchmarks and the average length of the videos (see creftypecap 0.A.2

). Moreover, even if the task itself is category-agnostic, empirical distribution of categories in the benchmarks tends to be heavily skewed towards a few common objects. We study whether this bias in category selection results in methods failing to generalize to more challenging objects by evaluating state-of-the-art user-initialized trackers on TAO in Section 

5.2. Semi-supervised video object segmentation differs from user-initialized tracking in that both the input to the tracker and the output are object masks, not boxes [46, 69]. As a result, such datasets are a lot more expensive to collect, and videos tend to be extremely short. The main focus of the works in this domain [12, 35, 61] is on accurate mask propagation, not solving challenging identity association problems, thus their effort is complementary to ours. Saliency-based tracking is an intriguing direction towards open-world tracking, where the objects of interest are defined not with a fixed vocabulary of categories, or manual annotations, but with bottom-up, motion- [45, 46] or appearance-based [13, 63] saliency cues. Our work similarly uses motion-based saliency to define a comprehensive vocabulary of categories, but presents a significantly larger benchmark with class labels for each object, enabling the use and evaluation of large-vocabulary object recognition approaches.

2.2 Algorithms

Multi-object trackers can be categorized into people and multi-category trackers. The former have been mainly developed on the MOT benchmark [42] and follow the tracking-by-detection paradigm, linking outputs of person detectors in an offline, graph-based framework [3, 4, 10, 20]. These methods mainly differ in the way they define the edge cost in the graph. Classical approaches use overlap between detections in consecutive frames [33, 47, 72]. More recent methods define edge costs based on appearance similarity [43, 51], or motion-based models [1, 15, 16, 37, 48, 54]. Very recently, Bergmann et al. [5] proposed a simple baseline approach that performs on par with SOTA people trackers, which repurposes an object detector’s bounding box regression capability to predict the position of an object in the next frame. Notice that all these methods have been developed and evaluated on the relatively small MOT dataset, which consists of 14 videos captured in very similar environments. By contrast, TAO provides a much richer, more diverse set of videos, encouraging trackers more robust to tracking challenges such as occlusion and camera motion. The more general multi-object tracking scenario is usually studied using ImageNet-Vid [52]. Methods in this group also use offline, graph-based optimization to link frame-level detections into tracks. To define the edge potentials, in addition to bounding box overlap, Feichtenhofer et al. [22] propose to use a similarity embedding, which is learned jointly with the detector. Alternatively, Kang et al. [34] directly predict short tubelets, and Xiao et al. [68] incorporate a spatio-temporal memory module inside a detector. Inspired by [5], we show that a simple baseline approach, relying on the Viterbi algorithm for linking detections across frames [22, 26], performs on par with the methods mentioned above on ImageNet-Vid. We then use this baseline for evaluating generic multi-object tracking on TAO in Section 5.2, and demonstrate that it struggles when faced with a large vocabulary and a diverse data distribution. User-initialized trackers tend to rely on a Siamese network architecture that was first introduced for signature verification [11], and later adapted for tracking [7, 18, 31, 57]. They learn a patch-level distance embedding and find the closest patch to the one annotated in the first frame in the following frames. To simplify the matching problem, state-of-the-art approaches limit the search space to the region in which the object was localized in the previous frame. Recently there have been several attempts to introduce some ideas from CNN architectures for object detection into Siamese trackers. In particular, Li et al. [39] use the similarity map obtained by matching the object template to the test frame as input to an RPN-like module adapted from Faster-RCNN [49]. Later this architecture was extended by introducing hard negative mining and template updating [76], as well as mask prediction [62]. In another line of work, Siamese-based trackers have been augmented with a target discrimination module to improve their robustness to distractors [9, 17]. We evaluate several state-of-the-art methods in this paradigm for which public implementation is available [9, 17, 18, 38, 62] on TAO, and demonstrate that they achieve only a moderate improvement over our multi-object tracking baseline, despite being provided with a ground truth initialization for each track (see Section 5.2 for details).

3 Dataset design

Our primary goal in this work is collecting a large-scale dataset of videos with a diverse vocabulary of labeled object tracks for evaluating trackers in the wild. This requires designing a strategy for (1) video collection, (2) vocabulary discovery, (3) scalable annotation, and (4) evaluation. We detail our strategies for (2-4) in this section, and defer the discussion of video collection to Section 4.1. Category discovery. Rather than manually defining a set of categories, we discover an object vocabulary from unlabeled videos which span diverse operating domains. Our goal is to focus on dynamic objects in the world. Towards this end, we ask annotators to mark all objects that move in our collection of videos, without any object vocabulary in mind. We then construct a vocabulary by giving names for all the discovered objects, following the recent trend for open-world dataset collection [28, 74]. In particular, annotators are asked to provide a free-form name for every object, but are encouraged to select a category from the LVIS [28] vocabulary whenever possible. We detail this process further in Section 4.2. Federation. Given this vocabulary, one option might be exhaustively labelling all instances of each category in all videos. Unfortunately, exhaustive annotation of a large vocabulary is expensive, even for images, as noted in [28]. We choose to use our labeling budget instead on collecting a large-scale, diverse dataset, by extending the federated annotation protocol of [28] from image datasets to videos. Rather than labeling every video with every category , we define three subsets of our dataset for each category: , which contains videos where all instances of are labeled, , videos with no instance of present in the video, and , videos where some instances of are annotated. Videos not belonging to any of these subsets are ignored when evaluating category . For each category , we only use videos in and to measure the precision of trackers, and videos in and to measure recall. We describe how to define , , and in Section 4.2. Granularity of annotations. To collect TAO, we choose to prioritize scale and diversity of the data at the cost of annotation granularity. In particular, we label tracks at 1 frame per second with bounding box labels but don’t annotated segmentation masks. This allows us to label 833 categories in 2,907 videos at a relatively modest cost. Our decision is motivated by the observation of [59] that dense frame labeling does not change the relative performance of the methods. Evaluation and metric. Traditionally, multi-object tracking datasets use either the CLEAR MOT metrics [6, 25, 42] or a 3D intersection-over-union (IoU) based metric [52, 70]. We report the former in creftypecap 0.B.2 (introducing modifications for large-vocabularies of classes, including multi-class aggregation and federation), but focus our experiments on the latter. To formally define 3D IoU, let and be a groundtruth and predicted track for a video with frames. 3D IoU is defined as: If an object is not present at time , we assign to an empty bounding box, and similarly for a missing detection. We choose 3D IoU (with a threshold of 0.5) as the default metric for TAO, and provide further analysis in creftypecap 0.B. Similar to standard object detection metrics, (3D) IoU together with (track) confidence can be used to compute mean average precision across categories. For methods that provide a score for each frame in a track, we use the average frame score as the track score. Following [28], we measure precision for a category in video only if all instances of the category are verified to be labeled in it.

4 Dataset collection

4.1 Video selection

Most video datasets focus on one or a few operating domains. For instance, MOT benchmarks [42] correspond to urban, outdoor scenes featuring crowds of people, whereas AVA [27] is sourced from produced films, typically capturing actors with close shots in carefully staged scenes. As a result, methods developed on any single dataset (and hence domain) fail to generalize in the wild. To avoid this bias, we constructed TAO by selecting videos from a variety of existing video benchmarks to ensure diversity of scenes and objects. Diversity. In particular, we used datasets for action recognition, self-driving cars, user-initialized tracking, as well as in-the-wild Flickr videos. In the action recognition domain we selected 3 datasets: Charades [56], AVA [27], and HACS [73]. Charades features complex human-human and human-object interactions, but all videos are indoor with limited camera motion. In contrast, AVA has a much wider variety of scenes and cinematographic styles but is scripted. HACS provides unscripted, in-the-wild videos. These action datasets are naturally focused on people and objects with which people interact. To include other animals and vehicles, we also source clips from LaSOT [21] (a benchmark for user-initialized tracking), BDD [71] and ArgoVerse [14] (benchmarks for self-driving cars). LaSOT is a diverse collection whereas BDD and ArgoVerse consist entirely of outdoor, urban scenes. Finally we sample in-the-wild videos from the YFCC100M [58] Flickr collection. Quality. The videos are automatically filtered to remove short videos and videos with a resolution below 480p. For longer videos, as in AVA, we use [41] to extract scenes without shot changes. In addition, we manually reviewed each sampled video to ensure it is high quality: i.e., we removed grainy videos as well as videos with excessive camera motion or shot changes. Finally, to focus on the most challenging tracking scenarios, we only kept videos that contain at least 2 moving objects. The full statistics of the collected videos are provided in Table 1. We point out that many prior video datasets tend to limit one or more quality dimensions (in terms of resolution, length, or number of videos) in order to keep evaluation and processing times manageable. In contrast, we believe that in order to truly enable tracking in the open-world, we need to appropriately scale benchmarks.

Figure 3: Our federated video annotation pipeline. First (a), annotators mine and track moving objects. Second (b), annotators categorize tracks using categories from the LVIS vocabulary or free-form text, producing the labeled tracks (c). Finally, annotators identify categories that are exhaustively annotated or verified to be absent. In this example (d), ‘person’s are identified as being exhaustively annotated, ‘camel’s are present but not exhaustively annotated and ‘bicycle’s and ‘mirror’s are verified as absent. Such federated labels allow one to accurately penalize false-positives and missed detections for exhaustively annotated and verified categories.

4.2 Annotation pipeline

Our annotation pipeline is illustrated in Figure 3. We designed it to separate low-level tracking from high-level semantic labeling. As pointed out by others [2], semantic labeling can be subtle and error-prone because of ambiguities and corner-cases that arise in category boundaries. By separating tasks into low vs high-level, we are able to take advantage of unskilled annotators for the former and highly-vetted workers for the latter.

Object mining and tracking.

We combine object mining and track labeling into a single stage of annotation. Given the set of videos described above, we ask annotators to mark objects that move at any point in the video. To avoid overspending our annotation budget on a few crowded videos, we limited the number of labeled object per video to 10. Note that this stage is category-agnostic: annotators are not instructed to look for objects from any specific vocabulary, but instead use motion as a saliency cue for mining relevant objects. They are then asked to track these objects throughout the video, and label them with bounding boxes at 1 frame-per-second. Finally, the tracks are verified by one independent annotator. This process is illustrated in Figure 3, where we can see that 6 objects are discovered and tracked.

Object categorization.

Next, we collected category labels for objects discovered in the previous stage and simultaneously constructed the dataset vocabulary. We focus on the large vocabulary from the LVIS [28] object detection dataset, which contains 1,230 synsets discovered in a bottom-up manner similar to ours. Doing so also allows us to make use of LVIS as a training set of relevant object detectors (which we later use within a tracking pipeline to produce strong baselines - creftypecap 5.1). Because maintaining a mental list of 1,230 categories is challenging even for expert annotators, we use an auto-complete annotation interface to suggests categories from the LVIS vocabulary (Figure 3 (b)). The autocomplete interface displays classes with a matching synset (e.g., “person.n.01”), name, synonym, and finally those with a matching definition. Interestingly, we find that some objects discovered in TAO, such as “door” or “marker cap”, do not exist in LVIS. To accommodate such important exceptions, we allow annotators to label objects with free-form text if they do not fit in the LVIS vocabulary. Overall, annotators labeled 16,144 objects (95%) with 488 LVIS categories, and 894 objects (5%) with 345 free-form categories. We use the 488 LVIS categories for MOT experiments (because detectors can be trained on LVIS), but use all categories for user-initialized tracking experiments in creftypecap 0.C.1.

Federated “exhaustive” labeling.

Finally, we ask annotators to verify which categories are exhaustively labeled for each video. Specifically, for each category labeled in video , we ask annotators whether all instances of are labeled. In creftypecap 3, after this stage, annotators marked that ‘person’ is exhaustively labeled, while ‘camel’ is not. Next, we show annotators a sampled subset of categories that are not labeled in the video, and ask them to indicate categories which are absent in the video. In creftypecap 3, annotators indicated that ‘bicycle’ and ‘mirror’ are absent.

4.3 Dataset splits

We intend for TAO to be used primarily as an evaluation benchmark. We split TAO into three subsets: train, validation and test, containing 500, 988 and 1,419 videos respectively. Typically, ‘train’ splits tend to be larger than ‘val’ and ‘test’. We choose to make TAO’s training set small for several reasons. Firstly, the primary goal of TAO is to reliably benchmark trackers in-the-wild. Secondly, most MOT systems are modularly trained using image-based detectors with hyper-parameter tuning of the overall tracking system. In our case, we ensure the train set is sufficiently large for hyper-parameter tuning, and ensure that our large-vocabulary is aligned with large-vocabulary image datasets (e.g., LVIS). This allows us to devote most of our annotation budget for large-scale ‘val’ and held-out ‘test’ sets.” We ensure that the videos in train, validation and test are well-separated. As an example, we ensure that each subject in the Charades dataset appears in only one of the train, validation or test sets. We provide further details on split construction in Appendix 0.A.3.

5 Analysis of state-of-the-art trackers

We now use TAO to analyze how well existing multi- and single-object trackers perform in the wild and when they fail. We tune the hyperparameters of each tracking approach on the ‘train’ set, and report results on the ‘val’ set. To capitalize on existing object detectors, we evaluate using the 488 LVIS categories in TAO. We begin by shortly describing the methods used in our analysis.

5.1 Methods

Detection. We analyze how well state-of-the-art object detectors perform on our dataset. To this end, we present results using a standard Mask R-CNN [49] detector trained using [66] in creftypecap 5.2.

Viterbi Det mAP Track mAP
Detection 73.4  [68] -
D&T [22] 79.8 -
STMN [68] 79.0 60.4
Detection 79.2 60.3
Table 2: ImageNet-Vid detection and track mAP; see text (left) for details.

Multi-Object Tracking. We analyze SOTA multi-object tracking methods on ImageNet-Vid, the largest vocabulary dataset prior to TAO. We first clarify whether such approaches improve detection or tracking. creftypecap 2 reports the standard ImageNet-Vid Detection mAP and Track mAP. The ‘Detection’ row corresponds to a detection-only baseline widely reported by prior work [68, 22, 75]. D&T [22] and STMN [68] are spatiotemporal architectures that produce SOTA improvements of 6-7% in detection mAP over a per-frame detector. However, both D&T and STMN post-process their per-frame outputs using the Viterbi algorithm, which iteratively links and re-weights the confidences of per-frame detections (see [26] for details). When the same post-processing is applied to a single-frame detector, one achieves nearly the same performance gain (Table 2, last row). Our analysis reinforces the bleak view of multi-object tracking progress suggested by [5]: while ever-more complex approaches have been proposed for the task, their improvements are often attributable to simple, baseline strategies. To foster meaningful progress on TAO, we evaluate a number of strong baselines in this work. We evaluate a powerful single-frame detector trained on LVIS [28] and COCO [40], followed by two linking methods: SORT [8], a simple, online linker initially proposed for tracking people, and the Viterbi post-processing step used by [22, 68], in creftypecap 5.2. Person detection and tracking. Detecting and tracking people have been a distinct focus in the multi-object tracking community. creftypecap 5.2 compares the above baselines to a recent SOTA people-tracker [5]. User-initialized tracking. We additionally present results using user-initialized trackers. We evaluate several recent methods for which public implementation is available [9, 17, 18, 38, 62]. Unfortunately, these trackers do not provide a class label for the objects they are tracking, and cannot directly be compared to multi-object trackers. However, these trackers can

be evaluated with an oracle classifier, allowing us to directly compare their accuracy with the methods that simultaneously detect and track objects.

Oracles. Finally, to disentangle the complexity of object classification and tracking, we use two oracles. The first, a class oracle, computes the best matching between predicted and groundtruth tracks in each video. Predicted tracks that match to a groundtruth track with 3D IoU are assigned the category of their matched groundtruth track. Tracks that do not match to a groundtruth track are not modified, and are treated as false positives. This allows us to evaluate the performance of trackers assuming the semantic classification task is solved. The second oracle computes the best possible assignment of per-frame detections to tracks, by comparing them with groundtruth. When doing so, class predictions for each detection are held constant. Any detections that are not matched are discarded. This oracle allows us to analyze the best performance we could expect given a fixed set of detections.

5.2 Results

How hard is object detection on TAO? We start by assessing the difficulty of the detection task on TAO. To this end we evaluate the SOTA object detector [29] using detection mAP. We train this model on a combination of LVIS and COCO, finding that training on LVIS alone led to a model that struggles to detect people. The final model achieves an mAP of 27.1 on TAO val at IoU 0.5, suggesting that single-frame detection is challenging on TAO.

Do multi-object trackers generalize to TAO?

creftypecap 5 reports results using tracking mAP on TAO. As a sanity check, we first evaluate a per-frame detector by assigning each detection to its own track. As expected, this achieves an mAP of nearly 0 (which isn’t quite 0 due to the presence of short tracks). Next, we evaluate two multi-object tracking approaches. We compare the SOTA Viterbi linking method to an online SORT tracker [8]. We tune SORT hyperparameters on our diverse ‘train’ set. Appendix 0.C.2 shows that this tuning is key for good accuracy. The offline Viterbi algorithm takes over a month of processing time to run on our ‘train’ set, prohibiting thorough parameter tuning. Instead, we tune a post-processing parameter for Viterbi: the score threshold for reporting a detection at each frame. We detail our tuning procedure in creftypecap 0.C.2. Surprisingly, we find that the simpler, online approach of SORT outperforms Viterbi, perhaps because the latter has been heavily tuned for ImageNet-Vid. Because of its scalablity (to many categories and long videos) and relatively better performance, we focus on SORT for the majority of our experiments. However, the performance of both methods remains low, suggesting TAO presents a major challenge for the tracking community, requiring principled novel approaches.

Oracle
Method Class Track Track mAP
Detection 0.6
Viterbi [22, 26] 6.3
SORT [8] 13.2
Detection 31.5
Viterbi [22, 26] 15.7
SORT [8] 30.2
Detection 83.6
Figure 4: SORT [8] and Viterbi linking [22, 26] provide strong baselines on TAO, but detection and tracking remain challenging. Relabeling and linking detections from current detectors using the class and track oracles is sufficient to achieve high performance, suggesting a pathway for progress on TAO.
Figure 5: SORT qualitative results, showing (left) a successful tracking result, and (right) a failure case due to semantic flicker between similar classes, suggesting that large-vocabulary tracking on TAO requires additional machinery.

To better understand the nature of the complexity of TAO, we separately measure the challenges of tracking and classification. To this end, we first evaluate the “track” oracle that perfectly links per-frame detections. It achieves a stronger mAP of 31.5, compared to 13.2 for SORT. Interestingly, providing SORT tracks with an oracle class label provides a similar improvement, boosting mAP to 30.2. We posit that these improvements are orthogonal, and verify this by combining them; we link detections with oracle tracks and assign these tracks oracle class labels. This provides the largest delta, dramatically improving mAP to 83.6%. This suggests that large-vocabulary tracking requires jointly improving tracking and classification accuracy (e.g., reducing semantic flicker as shown in Fig. 5).

Method Person AP
Viterbi [22, 26] 16.5
SORT [8] 18.5
Tracktor++ [5] 36.7
Table 3: Person-tracking results on TAO. See text (left) for details.

How well can we track people?

We now evaluate tracking on one particularly important category: people. Measuring AP for individual categories in a federated dataset can be noisy [28], so we emphasize relative performance of trackers rather than their absolute AP. We evaluate Tracktor++ [5], the state-of-the-art method designed specifically for people tracking on our dataset, and compare it to the SORT and Viterbi baselines in Table 3. For fairness, we update Tracktor++ to use the same detector used by our SORT and Viterbi baselines, but only use the ‘person’ predictions from this detector. Additionally, we tune the score threshold for Tracktor++ on our ‘train’ set, but find the method is largely robust to this parameter (see Appendix 0.C.2). We find that Tracktor++ strongly performs other approaches (36.7 AP), while SORT comes in second, modestly outperforming Viterbi (18.6 vs 16.5 AP). It is interesting to note that SORT, which can scale to all object categories, performs noticeably worse on all categories on average (13.2 mAP). creftypecap 0.B.2 shows that this delta between ‘person’ and other classes is even more dramatic using the MOTA metric (6.7 overall vs 54.8 for ‘person’). We attribute the higher accuracy for the ‘person’ category to two factors: (1) a rich history of focused research on this one category, which has led to more accurate detectors and trackers, and (2) more complex categories present significant challenges, such as hand-held objects which undergo repeated occlusions during interactions. To further investigate Tracktor++’s performance, we evaluate a simpler variant of the method from [5], which does not use appearance-based re-identification nor pixel-level frame alignment. We evaluate this variant on TAO, and find that removing these components reduces AP by over 8 points (from 36.7 to 25.9), suggesting that a majority of improvements over our baselines come from these two components. Our results contrast those of [5], which suggest that re-id and frame alignment are not particularly helpful. Compared to prior benchmarks, we posit the diversity of TAO results in a challenging testbed for person tracking which encourages trackers robust to occlusion and camera jitter.

Do user-initialized trackers generalize better?

Next, we present results of recent user-initialized trackers in creftypecap 4. For each object in TAO, we provide the user-initialized tracker with a groundtruth box. We consider two strategies for initialization. The standard approach (denoted ‘Init first’) initializes trackers using the first frame an object appears in, and runs trackers for the rest of the video. As the object may be partially occluded in this first frame, we additionally report a variant which initializes trackers using the frame with the largest bounding box (denoted ‘Init biggest’), and runs trackers forwards and backwards in time. Unlike multi-object trackers, most user-initialized trackers report a bounding box and confidence for objects at each frame, and do not explicitly report when an object is absent [59]. To resolve this, we modify each method to report an object as absent when the confidence drops below a threshold. We tune this threshold on the ‘train’ set in creftypecap 0.C.2 and find that user-initialized trackers are particularly sensitive to this threshold.

Oracle Track mAP
Method Box Init Class Init first Init biggest
SORT 30.2
ECO [18] 23.7 30.4
SiamMask [62] 30.8 37.0
SiamRPN++ LT [38] 27.2 30.4
SiamRPN++ [38] 29.7 35.9
ATOM* [17] 30.9 38.6
DIMP* [9] 33.2 38.5
Table 4: SOTA user-initialized tracking results on ‘val’. Surprisingly, despite using an oracle initial bounding box, these methods provide only modest improvements over a multi-object tracker. Because some user-initialized trackers are trained on videos in TAO, we re-train them on their original train set with TAO videos removed, denoting this with *.

We compare these trackers to SORT, supplying both with a class oracle. As expected, the use of a ground-truth initialization allows the best user-initialized methods to outperform the multi-object tracker. However, even with an oracle box initialization and an oracle classifier, tracking remains challenging on TAO. Indeed, most user-initialized trackers provide at most modest improvements over SORT, despite using an oracle box initialization. The ‘Init biggest’ strategy provides stronger improvements by initializing with easier frames, but this strategy cannot be used in online applications, as it requires access to the entire video. creftypecap 0.B.2 notes that user-initialized trackers can accurately track for a few frames after initialization, leading to improvements in MOTA, but provide little benefits in longer-term tracking. We hypothesize that the small improvement of user-initialized trackers over SORT is due to the fact that the former are trained on videos with a small vocabulary of objects with limited occlusions, leading to methods that do not generalize to the most challenging cases in TAO. One goal of user-initialized trackers is open-world tracking of objects without good detectors. TAO’s large vocabulary allows us to analyze progress towards this goal, indicating that large-vocabulary multi-object trackers may now address the open-world of objects as well as category-agnostic, user-initialized trackers.

6 Discussion

Developing tracking approaches that can be deployed in-the-wild requires being able to reliably measure their performance. With nearly 3,000 videos, TAO provides such a robust evaluation benchmark. Our analysis provides new conclusions about the state of tracking, while further raising a number of important questions to be explored in future work. The role of user-initialized tracking. User-initialized trackers aim to track any object, without requiring category-specific detectors. In this work, we raise a provocative question: with the advent of large vocabulary object detectors [28], to what extent can (detection-based) multi-object trackers perform generic tracking without user initialization? creftypecap 4, for example, shows that large-vocabulary datasets (such as TAO and LVIS) now allow multi-object trackers to match or outperform user-initialization for a number of categories.Specialized tracking approaches. Our hope in collecting TAO is to measure progress in tracking in-the-wild. A valid question is whether progress may be better achieved by building trackers for application-specific

scenarios. An indoor robot, for example, has little need for tracking elephants. However, success in many computer vision fields has been driven by the pursuit of

generic approaches, that can then be tailored for specific applications. We do not build one class of object detectors for indoor scenes, and another for outdoor scenes, and yet another for surveillance videos. We believe that tracking will similarly benefit from targeting diverse scenarios. Of course, due to its size, TAO also lends itself to use for evaluating trackers for specific scenarios or categories, as in creftypecap 5.2 for ‘person.’ Video object detection. Although image-based object detectors have shown significant improvements in recent years, our analysis in creftypecap 5.1 suggests that simple post-processing of detection outputs remains a strong baseline for detection in videos. While we do not emphasize it in this work, we note that TAO can also be used to measure progress in video object detection, where the goal is not to maintain the identity of objects, but simply to reliably detect them in each frame of a video. The large vocabulary in TAO particularly provides avenues for incorporating temporal information to resolve classification errors, which remain challenging (see creftypecap 5). Acknowledgements. We thank Jonathon Luiten and Ross Girshick for detailed feedback on the dataset and manuscript, and Nadine Chang and Kenneth Marino for reviewing early drafts. Annotations for this dataset were provided by Scale.ai. This work was supported in part by the CMU Argo AI Center for Autonomous Vehicle Research, the Inria associate team GAYA, and by the Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/Interior Business Center (DOI/IBC) contract number D17PC00345. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes not withstanding any copyright annotation theron. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied of IARPA, DOI/IBC or the U.S. Government.

Appendix

creftypecap 0.A further analyzes TAO annotations, including quality control and statistics. creftypecap 0.B further analyzes metrics, comparing 3D IoU to MOT challenge [42] metrics. Finally, creftypecap 0.C further analyzes tracking methods, providing results on non-LVIS categories and hyperparameter tuning experiments.

Appendix 0.A TAO annotations

This section presents additional details about TAO annotations. creftypecap 0.A.1 assesses the diversity and quality of annotations. creftypecap 0.A.2 analyzes the size, length and motion statistics of labeled tracks. Finally, creftypecap 0.A.3 provides further information regarding the construction of dataset splits.

0.a.1 Annotation diversity and quality

We analyze the diversity and quality of TAO annotations by re-annotating 50 videos in the dataset. Diversity. One might hope that this re-annotation closely matches the original annotation. However, in our federated setup, annotators are instructed to label only a subset of moving objects in each video. Thus, the annotations would only match if annotators had a bias towards a specific set of objects, which would hurt the diversity of TAO annotations. To verify whether this is the case, we check whether each track in the re-annotation corresponds to an object labeled in the original annotation. Concretely, if a re-annotated track has high overlap (IoU ) with a track in the original annotation, we assume the annotator is labeling the same object. Our re-annotation results in 310 tracks from 50 videos. Of these 310 tracks, just over half (177, or 57%) overlapped with those in the initial labeling with IoU . The rest were new objects not originally labeled in TAO, suggesting that annotators chose to label a diverse selection of objects. Quality. Next, we evaluate the annotation agreement of the 177 re-annotated tracks that correspond to tracks originally labeled in TAO. If our annotations are of high quality, we expect these tracks to have a very high IoU (say, ), as well as matching class labels. Indeed, the average IoU for the 177 overlapping tracks was 0.93, indicating annotators precisely labeled the spatial and temporal extent of objects. Finally, we evaluate the quality of the class labels in TAO. 165 (93%) were labeled with the same category as in the initial labeling; an additional 6 (3%) were labeled with a more precise or more general category (e.g., ‘jeep’ vs. ‘car’); finally, 6 were labeled with similar labels (e.g., ‘kayak’ vs. ‘canoe’) or other erroneous labels. This analysis indicates that despite the large vocabulary in TAO, the class labels in TAO are of high quality. If our annotations are of high quality, we expect these tracks to have a very high IoU (say, ), as well as matching class labels. Annotation details. We worked closely with a professional data-labeling company, Scale.ai, to label TAO. Each track was labeled by a Scale annotator, reviewed by Scale reviewers, and finally manually inspected by the authors.

0.a.2 Annotation statistics

We present further analysis of the annotated tracks in TAO in creftypecap 6. We compare TAO to MOT-17 [42] and ImageNet-Vid [52], which are benchmark datasets where the Viterbi [26, 22] and the Tracktor [5] approaches were originally evaluated. creftypecap 6(a) shows the distribution of changes in aspect ratio between two annotated frames at 1FPS. Concretely, the aspect ratio change is , where are the width and height of the object at time , respectively (see [36]). This metric can be used to understand the types of motion in tracking datasets. MOT-17 focuses on people, which largely have the same aspect ratio over time. ImageNet-Vid has a slightly more diverse distribution of changes in aspect ratio, but TAO has by far the most diverse distribution, due to its large size and diversity of categories. creftypecap 6(b) plots the distribution of bounding box resolution as a percentage of the image. MOT-17 tends to have smaller bounding boxes, while TAO and ImageNet-Vid have a variety of object sizes. Note again that TAO presents a much larger number of tracks used for evaluation, visible even on the log-scale in creftypecap 6(b), than ImageNet-Vid val. creftypecap 6(c) presents the distribution of object motion, proportional to the size of the object. Concretely, let be the area of the bounding box at time . We define the distance in as , and similarly for . Then, . As with creftypecap 6(a), we plot these changes at 1FPS so that the annotation rate does not impact the plot. We note that TAO contains a variety of object motions, including extremely fast motions for small objects, as evidenced by the number of boxes with motion change larger than 5.0. creftypecap 6(d) shows the distribution of object track lengths in TAO. For clarity, we group the tracks into 3 bins based on length: short, medium and long, which correspond to less than 1/3, between 1/3 and 2/3, and greater than 2/3 of the length of the video. The plot shows that TAO provides diversity in object track length, requiring methods to be able to track for long periods of time, while also being able to recognize when an object is missing. By contrast, MOT-17 is biased towards short tracks, while ImageNet-Vid is biased towards long tracks. Finally, we present statistics of four recent benchmarks for user-initialized tracking (or single-object tracking) in creftypecap 5. We note that datasets tend to benchmark tracking on a smaller number of categories than TAO, and on far fewer videos. While this may be appealing from a computational perspective, we argue that progress in tracking requires evaluating on a large, diverse set of scenarios, ensuring that methods do not overfit to any small set of videos or environments. Further, unlike standard user-initialized tracking datasets, TAO contains nearly 5x as many tracks per video, leading to a much larger number of total tracks compared to prior benchmarks.

(a) Ratio between aspect ratio of bounding boxes between two consecutive annotated frames at 1FPS.
(b) Bounding box size counts relative to size of the image.
(c) Distance between center of objects between two consecutive annotated frames at 1FPS.
(d) Track length counts, relative to video lengths.
Figure 6: Additional statistics of the TAO dataset. See creftypecap 0.A.2 for details.
Dataset
Classes
Eval. Train
Videos
Eval. Train
Avg
length (s)
Tracks
/ video
Min
resolution
Ann.
fps
Total Eval
length (s)
GOT-10k [32]222Stats from the GOT-10k dataset release, which differ from those in [32]. 84 480 360 9,335 12.2 1 270x480 10 4,384
OxUvA [59] 22 0 366 0 141.2 1.1 192x144 1 51,667
LaSOT [21] 70 70 280 1,120 82.1 1 202x360 ~25 23,520
TrackingNet [44] 27 27 511 30,132 14.7 1 270x360 ~28 7,511
TAO (Ours)333TAO train and eval contain partially overlapping subsets of the overall 833 categories. 785 316 2,407 500 36.8 5.9 640x480 1 88,605
Table 5: Statistics of major user-initialized tracking datasets.

0.a.3 Split construction

We construct our ‘train’, ‘val’, and ‘test’ splits to respect the following constraints:

  • Charades contains videos recorded by mechanical turk workers, and one worker may contribute multiple videos to Charades. We ensure that any two videos uploaded by the same worker falls in the same split.

  • ArgoVerse contains video recordings from different cameras from the same driving sequence. We ensure that all videos from the same driving sequence fall in the same split.

  • HACS contains videos uploaded to YouTube. Any two videos uploaded by the same YouTube user, or uploaded to the same YouTube channel, must fall in the same split.

  • AVA. We split AVA movies into multiple contiguous shots, and ensure shots from the same movie fall in the same split.

  • YFCC100M contains videos uploaded to Flickr. Any two videos uploaded by the same Flickr user fall in the same split.

  • BDD and LaSOT: No constraints are applied for split construction.

Appendix 0.B Metrics

In this section, we further analyze the 3D IoU metric (0.B.1), report results using the MOT challenge [42] metrics (0.B.2), and finally present per-category APs for SORT (0.B.3).

0.b.1 3D IoU Discussion

The mAP metric using 3D IoU provides a concise, interpretable evaluation of tracking in the wild, as evidenced by its use in recent datasets for multi-object tracking with many categories [19, 70]. We further discuss this metric below: Relation to identity swaps. creftypecap 7 shows that 3D IoU is correlated with a key metric for tracking: identity swaps, as measured by the MOT challenge [42] metrics.

Figure 7:

For each pair of predicted and groundtruth tracks matched to each other on TAO, we compute the 3D IoU and number of ID swaps. Above, we plot the mean and variance of 3D IoU vs. ID swaps across tracks, and show that 3D IoU drops as the number of ID swaps increases.

Partial credit. Evaluating trackers with mAP requires specifying an IoU threshold, which we set to throughout the experiments in the main paper. Consequentially, trackers do not receive partial credit for tracking an object for short time periods. Consider two trackers: Tracker A perfectly tracks an object for 30% of its track length, while Tracker B only tracks the object 5% of the time. At an IoU threshold of 0.5, A and B will result in the same mAP. By contrast, metrics such as MOTA and ID-F1 will be significantly higher for A than for B. The 3D IoU mAP metric takes inspiration from image-based detection metrics: as object detectors receive no credit for loose localizations, object trackers receive no credit for loosely tracking objects for a few frames. If desired, the mAP metric can be modified to provide partial credit by averaging over multiple IoU thresholds, similar to the COCO evaluation [40].

Confidence estimates.

Metrics such as MOTA [6] and ID-F1 [65]

metrics do not evaluate the confidence provided by many modern tracking approaches. By contrast, our mAP metric evaluates these explicitly when tracing out the precision-recall curve. This allows us to evaluate methods across diverse application scenarios, which may have different tradeoffs between precision and recall.

Impact of object size. 3D IoU is computed over spatio-temporal volumes. As such, frames where an object’s bounding box is large have a greater impact on the spatio-temporal volume than frames where an object’s bounding box is small, thus factoring in more heavily into the IoU measure. We note that for many applications, such as navigation, this is a desirable property, as accurate localization and tracking is more important for nearby objects. For other applications, additional diagnostics, such as MOTA (creftypecap 0.B.2), can be used for further analysis.

0.b.2 MOTA results

For completeness, we present results using the MOT challenge suite of metrics [42]: MOTA [6], ID-F1 [50], mostly-tracked (MT) tracks and mostly-lost (ML) tracks [65], false-positives (FP), false-negatives (FN) and identity swaps (ID Sw.), computed using the py-motmetrics library [30]. To do this, we first make two modifications to the MOT metrics: Federated MOTA and ID-F1. We update the MOTA and ID-F1 metrics for a federated dataset by only counting false positives (FPs) for a category in video if we know that all instances of category are annotated in video (i.e., if is in or as defined in Sec. 3 of our paper). While this approach is not perfect, as it can over-estimate the performance of a tracker, it provides a simple adaptation to the federated setup. Multiple categories. The MOT metrics are usually reported for a single category [42], or separately for a small number categories [24]. This is not a scalable strategy for TAO, which contains 833 categories. Instead, we compute metrics separately per category, and combine them across categories. Concretely, for metrics such as MOTA and ID-F1, we report the average value across categories. For counters, including MT (mostly-tracked), ML (mostly-lost), FP (false-positives), FN (false-negatives) and ID Sw. (identity switches), we report the sum across categories. Note that while MOTA and ID-F1 are balanced across categories, the ‘counters’ are heavily dominated by the most frequent categories. Thresholds. Unlike mAP, the MOT metrics require picking a confidence threshold for evaluation. To do this, we search over track score thresholds on TAO train and report results in creftypecap 6. For Viterbi and user-initialized trackers, the track score threshold is applied after the tracker per-frame score threshold tuned in creftypecap 0.C.2. Hence, the MOTA for track thresholds below the per-frame threshold are equivalent (e.g., for DIMP, the optimal per-frame threshold is 0.5, and so the MOTA for thresholds below 0.5 is exactly the same: 22.7).

Tracker 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Detector -18.1 -9.7 -6.1 -3.8 -2.2 -1.3 -1.0 -0.3 -0.01 0.0
SORT -3.0 7.7 7.7 7.9 8.5 6.9 5.4 3.6 2.5 0.0
Viterbi -8.4 2.5 5.4 5.6 6.2 6.8 5.3 5.3 3.3 0.0
ATOM 21.8 21.8 21.8 21.8 21.8 21.8 27.2 19.8 8.2 0.0
DIMP 22.7 22.7 22.7 22.7 22.7 22.7 22.6 21.4 20.3 19.1
ECO 0.7 0.7 0.3 1.5 7.0 8.1 12.6 6.1 0.3 0.0
SiamMask 19.7 19.7 19.7 19.7 19.7 19.7 19.7 19.7 19.9 0.0
SiamRPN++ 21.0 21.0 21.0 21.0 21.0 21.0 21.0 21.0 25.5 0.0
SiamRPN++ LT 22.9 22.9 22.9 22.9 22.9 22.9 22.9 22.9 22.9 0.0
Person-only evaluation
Tracktor++ 65.9 66.0 66.2 66.5 67.2 67.9 68.4 67.8 63.0
Table 6: Results from tuning track score thresholds for multi-object trackers, user-initialized trackers, and Tracktor++ on TAO train, reporting MOTA.

We use the optimal thresholds from the train set to report results on the validation set for multi-object trackers in creftypecap 7, for user-initialized trackers in creftypecap 8, and for person-tracking in creftypecap 9. In general, we find that the conclusions drawn in our main paper using mAP are consistent with experiments using MOTA, with two exceptions.

Oracle
Method Class Track MOTA ID-F1 MT ML FP FN ID Sw.
Detection -2.3 1.3 1,495 1,941 3,492 60,776 48,377
Viterbi 5.6 10.0 1,407 2,409 5,367 62,341 10,262
SORT 6.7 10.4 1,687 2,117 4,146 59,481 4,772
Detection 38.8 48.4 2,191 919 0 42,796 0
Viterbi 8.3 13.8 1447 2361 5595 60787 10292
SORT 11.3 15.6 1,725 2,066 4,165 58,418 4,773
Detection 83.2 89.6 3,806 188 0 17018 6
Table 7: MOT challenge metrics for multi-object trackers on TAO validation. As the ‘Track’ oracle implicitly removes false positive detections, we set score thresholds to 0 when it is used.

User init.

First, creftypecap 8 shows that user-initialized trackers provide significant improvements over SORT using MOTA and ID-F1, while this did not hold for mAP. These metrics provide partial credit for tracking objects for short periods of time, while mAP (with an 3D IoU threshold of 0.5) requires tracking an object for at least half its track length (see creftypecap 0.B.1). One can obtain mAP rankings consistent with MOTA/ID-F1 by using an artificially low IoU threshold; at a threshold of 0.1, DIMP strongly outperforms SORT, 71.0 mAP to 36.9 mAP. These results reinforce the notion that user-initialization is helpful for tracking short periods after initialization, but less helpful in the long term.

Oracle
Method Box Init Class MOTA ID-F1 MT ML FP FN ID Sw.
SORT 11.3 15.6 1,725 2,066 4,165 58,418 4,773
ECO 11.8 24.0 753 4341 5395 85415 42
SiamRPN++ LT 13.1 54.0 2,292 753 19282 42255 2103
SiamRPN++ 14.6 49.9 2,110 1229 16630 45612 1411
ATOM 16.9 46.7 1,694 2,274 14,625 55,875 481
DIMP 24.4 55.1 2,279 870 16,966 42,729 1,290
Table 8: MOT challenge metrics on TAO validation, comparing user-initialized trackers with SORT using a class oracle.

MOTA-Person.

Second, as noted in the main paper, creftypecap 9 shows that MOTA-person is significantly higher than MOTA-overall (6.7 vs 54.8 for SORT), whereas the delta is smaller under mAP (13.2 vs 18.5 for SORT). We find MOT metrics heavily reward accurate detection while 3D IoU heavily penalizes inaccurate tracking. Because person detectors strongly outperform other category detectors on average, this is manifested as a high MOTA-person score.

Method MOTA ID-F1 MT ML FP FN ID Sw.
Viterbi 44.5 50.4 939 741 21,678 3,167 7,128
SORT 54.8 56.2 1,078 542 20,025 2,432 3,567
Tracktor++ 66.6 64.8 1,529 411 12,910 2,821 3,487
Table 9: MOT challenge metrics on TAO validation for the ‘person’ category.

Other benchmarks. Finally, we directly compare Tracktor++ on TAO with its performance on the MOT-17 dataset. creftypecap 10 shows that the more sophisticated components of Tracktor++ (re-identification and motion compensation) lead to significant improvements on TAO, suggesting TAO encourages trackers robust to common tracking challenges, including occlusion and camera motion.

TAO MOT-17
Method train val train test
Tracktor 63.8 61.6 61.5 -
Tracktor++ (reID + CMC) 68.4 66.6 61.9 53.5
Table 10: MOTA on TAO val vs. MOT-17, for Tracktor. TAO encourages trackers robust to camera motion and occlusion, as noted by the significant improvement to Tracktor using the reID and camera motion compensation (CMC) components.

0.b.3 AP per category

We present per-category APs in creftypecap 8 for the SORT algorithm reported in the main paper, though we note that AP for individual categories can be noisy in a federated setup [28]. Note that for 180 categories, this algorithm achieves 0 AP; for conciseness, we plot only the categories with non-zero AP.

Figure 8: Per-category AP for the SORT algorithm, omitting 180 categories which result in zero AP for conciseness. As common in large-vocabulary datasets (LVIS, ADE-20K, LabelMe), average accuracy is dominated by classes in the tail, many of which result in 0 AP. Note that AP for individual categories can be noisy in a federated setup [28].

Appendix 0.C Additional tracking results

creftypecap 0.C.1 presents results for user-initialized trackers on all categories in TAO. creftypecap 0.C.2 reports results from tuning trackers on TAO train.

0.c.1 User-initialized trackers on all categories

In the main paper, we focus our analysis on a subset of TAO categories which exist in the LVIS [28] dataset, allowing us to repurpose existing object detectors for multi-object tracking. Here, we evaluate user-initialized trackers (which do not require object detectors) on the remaining categories in TAO (creftypecap 11). We generally find that the results are consistent with the results on the LVIS categories.

Method Non-LVIS categories, validation
ECO 24.1
SiamMask 27.0
SiamRPN++ 27.7
SiamRPN++ LT 25.1
ATOM 29.5
DIMP 29.6
Table 11: Results on non-LVIS, free-form text categories in TAO validation.

0.c.2 Hyperparameter tuning

This section reports detailed results of tuning each tracker on TAO train, as well as information about the detector used for SORT, Viterbi and Tracktor++ (0.C.3). Preliminary: Score thresholds. Before discussing the details of each tracker, we define three different score thresholds used by trackers, and refer to them by name throughout the appendix:

  1. Detection score: This is the confidence reported by a detector for each object at each frame, before any tracking has taken place.

  2. Tracker per-frame score: This is the confidence reported by the tracker for each object at each frame, after tracking is complete.

  3. Track score: This is the confidence reported by the tracker for each object track throughout the video. This confidence is used to rank tracks when computing mAP. When computing MOTA, we tune the threshold for reporting tracks using the track score, as described in creftypecap 0.B.2.

Sort.

We tune three parameters internal to SORT, as well as parameters of the underlying detector in creftypecap 12. We tune the following SORT parameters:

  1. Det / image: Max number of detections output by the detector per image.

  2. Detection score

  3. max_age: How many frames tracks are kept ‘alive’ for, without any detections being matched to them.

  4. min_hits: How many frames a track must be alive for before it is considered ‘confirmed’ and output.

  5. min_iou: Minimum IoU between a track and a detection required for linking the two.

  6. NMS Thresh: The NMS IoU threshold used by the detector. We experiment with more aggressive NMS, which may make the task of linking detections using IoU easier.

The first row in creftypecap 12 corresponds to the default SORT parameters. Due to the significant motion and long duration of sequences in TAO (see creftypecap 0.A.2), we find that increasing max_age and decreasing min_iou and min_hits helps significantly with accuracy. Additionally, we find that outputting more boxes per image consistently improves accuracy. Lowering the score threshold from 0.1 to 0.0005 results in a 2.1 point improvement from 8.2 to 11.3, and lowering the NMS and score thresholds provides even more significant improvements, from 11.3 to 16.3.

Params
NMS Thresh Det / image Det score max_age min_hits min_iou Track mAP
0.5 300 0.1 1 3 0.3 4.3
0.5 300 0.1 1 3 0.1 5.0
0.5 300 0.1 1 3 0.5 4.3
0.5 300 0.1 1 1 0.1 5.1
0.5 300 0.1 1 5 0.1 4.9
0.5 300 0.1 1 10 0.1 4.9
0.5 300 0.1 10 1 0.1 6.5
0.5 300 0.1 50 1 0.1 8.1
0.5 300 0.1 100 1 0.1 8.2
0.5 300 0.001 100 1 0.1 10.5
0.5 300 0.0005 100 1 0.1 11.3
0.5 300 0.0001 100 1 0.1 10.9
0.5 10,000 0.0005 100 1 0.1 9.4
0.1 10,000 0.0005 100 1 0.1 15.3
0 10,000 0.0005 100 1 0.1 16.3
Table 12: Results from tuning SORT parameters (by coordinate descent) on TAO train, where the active coordinate (parameter) is highlighted.

Viterbi.

The Viterbi approach has a number of tunable parameters. Unfortunately, the code for this approach is prohibitively expensive to run, taking over a week of compute time to process TAO train in parallel on 4 machines. Due to this constraint, we do not tune the internal parameters of this approach. However, creftypecap 13 shows that tuning the tracker’s per-frame score post-hoc can provide small improvements in accuracy, from 8.5 to 9.0.

Tracker per-frame score 0 0.1 0.2 0.3 0.4 0.5
Track mAP 8.5 9.0 8.4 8.4 7.8 7.3
Table 13: Results from tuning the Viterbi tracker’s per-frame score threshold TAO train.

Tracktor++.

Tracktor++ by default thresholds the output of a detector at 0.5. creftypecap 14 shows the results of tuning this threshold on TAO train. Perhaps surprisingly, we find that Tracktor++ is fairly robust to this parameter, unlike SORT (as seen in creftypecap 12). We hypothesize that this may be because of two Tracktor++ components: (1) the use of detections at time as proposal at time may make detectors more likely to consistently output high-confidence detections for tracks, and (2) the re-id component may allow Tracktor++ to more accurately recover tracks with no matching detections for a few frames.

Detection score 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Track mAP 35.1 35.5 35.5 35.7 35.0 34.7 34.6 33.0 29.8
Table 14: Results from tuning Tracktor’s detection score threshold on TAO’s train set.

User-initialized trackers.

As user-initialized trackers do not explicitly report when an object is absent, we modify each method to report an object as absent when the confidence drops below a threshold. We tune this threshold on TAO train. creftypecap 15 shows that the optimal threshold varies by tracker, and tuning this parameter can lead to significant changes in accuracy (e.g., 5.2% in the case of DIMP when using a threshold of 0.5 as opposed to the default of 0).

Tracker Tracker per-frame score
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.99
ATOM 34.3 34.3 36.2 36.6 36.7 34.4 31.9 26.3 17.8 9.9 2.1
DIMP 31.2 33.2 35.3 36.1 36.2 36.4 34.8 33.2 30.3 27.8 19.7
ECO 25.4 25.4 26.3 27.3 27.1 25.6 20.6 14.7 8.9 3.0 2.7
SiamMask 27.9 28.6 28.8 28.8 29.3 29.4 30.0 30.7 30.9 30.5 27.3
SiamRPN++ 28.6 29.2 29.3 30.3 30.0 30.9 31.1 31.5 31.2 31.4 28.1
SiamRPN++-LT 27.0 26.6 27.1 27.2 27.0 27.7 28.0 27.9 28.0 28.2 26.7
Table 15: Results from tuning user-initialized trackers’ per-frame score threshold on TAO train.
Figure 9: Qualitative comparison between a Mask R-CNN model trained on LVIS (left) and one trained on LVIS+COCO (right). Training on additional COCO data is critical for accurately detecting common categories, such as people and cars.

0.c.3 Detector details

Throughout our experiments, we used a Mask R-CNN model [29] using a ResNet-101 backbone. We re-train this model on a combination of the LVIS and COCO datasets (described below) using the default training parameters for training on LVIS (including repeat factor sampling). Specifically, we used the detectron2 repository [66], with the configuration file at https://github.com/facebookresearch/detectron2/blob/b6fe828a2f3b2133f24cb93c1d0d74cb59c6a15d/configs/LVIS-InstanceSegmentation/mask_rcnn_R_101_FPN_1x.yaml. We found that training on a combination of COCO and LVIS annotations leads to a noticeable improvement in detection quality, which is particularly significant for people, compared to training on LVIS alone. To build this combination, we add COCO annotations to every image in the LVIS dataset. To avoid duplicates, we remove COCO annotations that have IoU with an LVIS annotation. We show qualitative results of this improvement in creftypecap 9.

References

  • [1] Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, and Silvio Savarese. Social LSTM: Human trajectory prediction in crowded spaces. In CVPR, 2016.
  • [2] Adela Barriuso and Antonio Torralba. Notes on image annotation. arXiv preprint arXiv:1210.3448, 2012.
  • [3] Jerome Berclaz, Francois Fleuret, and Pascal Fua. Robust people tracking with global trajectory optimization. In CVPR, 2006.
  • [4] Jerome Berclaz, Francois Fleuret, Engin Turetken, and Pascal Fua. Multiple object tracking using k-shortest paths optimization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(9):1806–1819, 2011.
  • [5] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In ICCV, 2019.
  • [6] Keni Bernardin and Rainer Stiefelhagen. Evaluating multiple object tracking performance: the clear mot metrics. Journal on Image and Video Processing, 2008:1, 2008.
  • [7] Luca Bertinetto, Jack Valmadre, Joao F Henriques, Andrea Vedaldi, and Philip HS Torr. Fully-convolutional Siamese networks for object tracking. In ECCV, 2016.
  • [8] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In ICIP, 2016.
  • [9] Goutam Bhat, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning discriminative model prediction for tracking. In CVPR, 2019.
  • [10] Michael D Breitenstein, Fabian Reichlin, Bastian Leibe, Esther Koller-Meier, and Luc Van Gool. Robust tracking-by-detection using a detector confidence particle filter. In ICCV, 2009.
  • [11] Jane Bromley, Isabelle Guyon, Yann LeCun, Eduard Säckinger, and Roopak Shah.

    Signature verification using a ”Siamese” time delay neural network.

    In NIPS, 1994.
  • [12] Sergi Caelles, Kevis-Kokitsi Maninis, Jordi Pont-Tuset, Laura Leal-Taixé, Daniel Cremers, and Luc Van Gool. One-shot video object segmentation. In CVPR, 2017.
  • [13] Sergi Caelles, Jordi Pont-Tuset, Federico Perazzi, Alberto Montes, Kevis-Kokitsi Maninis, and Luc Van Gool. The 2019 DAVIS challenge on VOS: Unsupervised multi-object segmentation. arXiv preprint arXiv:1905.00737, 2019.
  • [14] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In CVPR, 2019.
  • [15] Boyu Chen, Dong Wang, Peixia Li, Shuang Wang, and Huchuan Lu. Real-time ’Actor-Critic’ tracking. In ECCV, 2018.
  • [16] Wongun Choi and Silvio Savarese. Multiple target tracking in world coordinate with single, minimally calibrated camera. In ECCV, 2010.
  • [17] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. Atom: Accurate tracking by overlap maximization. In CVPR, 2019.
  • [18] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. Eco: Efficient convolution operators for tracking. In CVPR, 2017.
  • [19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009.
  • [20] Andreas Ess, Bastian Leibe, Konrad Schindler, and Luc Van Gool. A mobile vision system for robust multi-person tracking. In CVPR, 2008.
  • [21] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. LaSOT: A high-quality benchmark for large-scale single object tracking. In CVPR, 2019.
  • [22] Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Detect to track and track to detect. In ICCV, 2017.
  • [23] Robert Fisher, J Santos-Victor, and J Crowley. Context aware vision using image-based active recognition. EC’s Information Society Technology’s Programme Project IST2001-3754, 2001.
  • [24] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The KITTI dataset. The International Journal of Robotics Research, 32(11):1231–1237, 2013.
  • [25] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? The KITTI vision benchmark suite. In CVPR, 2012.
  • [26] Georgia Gkioxari and Jitendra Malik. Finding action tubes. In CVPR, 2015.
  • [27] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. AVA: A video dataset of spatio-temporally localized atomic visual actions. In CVPR, 2018.
  • [28] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In CVPR, 2019.
  • [29] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask R-CNN. In ICCV, 2017.
  • [30] Christoph Heindl and Jack Valmadre. py-motmetrics. https://github.com/cheind/py-motmetrics/, 2019.
  • [31] David Held, Sebastian Thrun, and Silvio Savarese. Learning to track at 100 FPS with deep regression networks. In ECCV, 2016.
  • [32] Lianghua Huang, Xin Zhao, and Kaiqi Huang. GOT-10k: A large high-diversity benchmark for generic object tracking in the wild. arXiv preprint arXiv:1810.11981, 2018.
  • [33] Hao Jiang, Sidney Fels, and James J Little.

    A linear programming approach for multiple object tracking.

    In CVPR, 2007.
  • [34] Kai Kang, Hongsheng Li, Tong Xiao, Wanli Ouyang, Junjie Yan, Xihui Liu, and Xiaogang Wang. Object detection in videos with tubelet proposal networks. In CVPR, 2017.
  • [35] Anna Khoreva, Rodrigo Benenson, Eddy Ilg, Thomas Brox, and Bernt Schiele. Lucid data dreaming for video object segmentation. International Journal of Computer Vision, 127(9):1175–1197, 2019.
  • [36] Matej Kristan, Jiri Matas, Aleš Leonardis, Tomáš Vojíř, Roman Pflugfelder, Gustavo Fernandez, Georg Nebehay, Fatih Porikli, and Luka Čehovin. A novel performance evaluation methodology for single-target trackers. TPAMI, 38(11):2137–2155, 2016.
  • [37] Laura Leal-Taixé, Michele Fenzi, Alina Kuznetsova, Bodo Rosenhahn, and Silvio Savarese. Learning an image-based motion context for multiple people tracking. In CVPR, 2014.
  • [38] Bo Li, Wei Wu, Qiang Wang, Fangyi Zhang, Junliang Xing, and Junjie Yan. Siamrpn++: Evolution of siamese visual tracking with very deep networks. In CVPR, 2019.
  • [39] Bo Li, Junjie Yan, Wei Wu, Zheng Zhu, and Xiaolin Hu. High performance visual tracking with siamese region proposal network. In CVPR, 2018.
  • [40] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014.
  • [41] Jakub Lokoč, Gregor Kovalčík, Tomás Souček, Jaroslav Moravec, and Premysl Čech. A framework for effective known-item search in video. In ACMM, 2019.
  • [42] Anton Milan, Laura Leal-Taixé, Ian Reid, Stefan Roth, and Konrad Schindler. MOT16: A benchmark for multi-object tracking. arXiv preprint arXiv:1603.00831, 2016.
  • [43] Anton Milan, S Hamid Rezatofighi, Anthony Dick, Ian Reid, and Konrad Schindler.

    Online multi-target tracking using recurrent neural networks.

    In

    Thirty-First AAAI Conference on Artificial Intelligence

    , 2017.
  • [44] Matthias Muller, Adel Bibi, Silvio Giancola, Salman Alsubaihi, and Bernard Ghanem. Trackingnet: A large-scale dataset and benchmark for object tracking in the wild. In ECCV, 2018.
  • [45] Peter Ochs, Jitendra Malik, and Thomas Brox. Segmentation of moving objects by long term video analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(6):1187–1200, 2013.
  • [46] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, 2016.
  • [47] Hamed Pirsiavash, Deva Ramanan, and Charless C Fowlkes. Globally-optimal greedy algorithms for tracking a variable number of objects. In CVPR, 2011.
  • [48] Liangliang Ren, Jiwen Lu, Zifeng Wang, Qi Tian, and Jie Zhou.

    Collaborative deep reinforcement learning for multi-object tracking.

    In ECCV, 2018.
  • [49] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NIPS, 2015.
  • [50] Ergys Ristani, Francesco Solera, Roger Zou, Rita Cucchiara, and Carlo Tomasi. Performance measures and a data set for multi-target, multi-camera tracking. In ECCV, 2016.
  • [51] Ergys Ristani and Carlo Tomasi. Features for multi-target multi-camera tracking and re-identification. In CVPR, 2018.
  • [52] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. ImageNet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211–252, 2015.
  • [53] Bryan C Russell, Antonio Torralba, Kevin P Murphy, and William T Freeman. LabelMe: a database and web-based tool for image annotation. International journal of computer vision, 77(1-3):157–173, 2008.
  • [54] Paul Scovanner and Marshall F Tappen. Learning pedestrian dynamics from the real world. In ICCV, 2009.
  • [55] Xindi Shang, Donglin Di, Junbin Xiao, Yu Cao, Xun Yang, and Tat-Seng Chua. Annotating objects and relations in user-generated videos. In ICMR, 2019.
  • [56] Gunnar A Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In ECCV, 2016.
  • [57] Ran Tao, Efstratios Gavves, and Arnold WM Smeulders. Siamese instance search for tracking. In CVPR, 2016.
  • [58] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. arXiv preprint arXiv:1503.01817, 2015.
  • [59] Jack Valmadre, Luca Bertinetto, Joao F Henriques, Ran Tao, Andrea Vedaldi, Arnold WM Smeulders, Philip HS Torr, and Efstratios Gavves. Long-term tracking in the wild: A benchmark. In ECCV, 2018.
  • [60] Paul Voigtlaender, Michael Krause, Aljosa Osep, Jonathon Luiten, Berin Balachandar Gnana Sekar, Andreas Geiger, and Bastian Leibe. MOTS: Multi-object tracking and segmentation. In CPVR, 2019.
  • [61] Paul Voigtlaender and Bastian Leibe.

    Online adaptation of convolutional neural networks for video object segmentation.

    In BMVC, 2017.
  • [62] Qiang Wang, Li Zhang, Luca Bertinetto, Weiming Hu, and Philip HS Torr. Fast online object tracking and segmentation: A unifying approach. In CVPR, 2019.
  • [63] Wenguan Wang, Hongmei Song, Shuyang Zhao, Jianbing Shen, Sanyuan Zhao, Steven CH Hoi, and Haibin Ling. Learning unsupervised video object segmentation through visual attention. In CVPR, 2019.
  • [64] Longyin Wen, Dawei Du, Zhaowei Cai, Zhen Lei, Ming-Ching Chang, Honggang Qi, Jongwoo Lim, Ming-Hsuan Yang, and Siwei Lyu. UA-DETRAC: A new benchmark and protocol for multi-object detection and tracking. arXiv preprint arXiv:1511.04136, 2015.
  • [65] Bo Wu and Ram Nevatia. Tracking of multiple, partially occluded humans based on static body part detection. In CVPR, 2006.
  • [66] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detectron2. https://github.com/facebookresearch/detectron2, 2019.
  • [67] Yi Wu, Jongwoo Lim, and Ming-Hsuan Yang. Online object tracking: A benchmark. In CVPR, 2013.
  • [68] Fanyi Xiao and Yong Jae Lee. Video object detection with an aligned spatial-temporal memory. In ECCV, 2018.
  • [69] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-VOS: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018.
  • [70] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019.
  • [71] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, June 2020.
  • [72] Li Zhang, Yuan Li, and Ramakant Nevatia. Global data association for multi-object tracking using network flows. In CVPR, 2008.
  • [73] Hang Zhao, Antonio Torralba, Lorenzo Torresani, and Zhicheng Yan. HACS: Human action clips and segments dataset for recognition and temporal localization. In ICCV, 2019.
  • [74] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ADE20k dataset. In CVPR, 2017.
  • [75] Xizhou Zhu, Yujie Wang, Jifeng Dai, Lu Yuan, and Yichen Wei. Flow-guided feature aggregation for video object detection. 2017 IEEE International Conference on Computer Vision (ICCV), pages 408–417, 2017.
  • [76] Zheng Zhu, Qiang Wang, Bo Li, Wei Wu, Junjie Yan, and Weiming Hu. Distractor-aware siamese networks for visual object tracking. In ECCV, 2018.