Object detection has received great attention during recent years. Pedestrian detection is a canonical sub-problem that remains a popular topic of research due to its diverse applications.
Despite the extensive research on pedestrian detection, recent papers still show significant improvements, suggesting that a saturation point has not yet been reached. In this paper we analyse the gap between the state of the art and a newly created human baseline (section 3.1). The results indicate that there is still a ten fold improvement to be made before reaching human performance. We aim to investigate which factors will help close this gap.
We analyse failure cases of top performing pedestrian detectors and diagnose what should be changed to further push performance. We show several different analysis, including human inspection, automated analysis of problem cases (e.g. blur, contrast), and oracle experiments (section 3.2). Our results indicate that localisation is an important source of high confidence false positives. We address this aspect by improving the training set alignment quality, both by manually sanitising the Caltech training annotations and via algorithmic means for the remaining training samples (sections 3.3 and 4.1).
To address background versus foreground discrimination, we study convnets for pedestrian detection, and discuss which factors affect their performance (section 4.2).
1.1 Related work
In the last years, diverse efforts have been made to improve the performance of pedestrian detection. Following the success of integral channel feature detector (ICF) [6, 5], many variants [22, 24, 16, 18, 23] were proposed and showed significant improvement. A recent review of pedestrian detection  concludes that improved features have been driving performance and are likely to continue doing so. It also shows that optical flow  and context information  are complementary to image features and can further boost detection accuracy.
By fine-tuning a model pre-trained on external data convolution neural networks (convnets) have also reached state-of-the-art performance[15, 20].
Most of the recent papers focus on introducing novelty and better results, but neglect the analysis of the resulting system. Some analysis work can be found for general object detection [1, 14]; in contrast, in the field of pedestrian detection, this kind of analysis is rarely done. In 2008,  provided a failure analysis on the INRIA dataset, which is relatively small. The best method considered in the 2012 Caltech dataset survey  had more false positives at recall than the methods considered here, and no method had reached the mark.
Since pedestrian detection has improved significantly in recent years, a deeper and more comprehensive analysis based on state-of-the-art detectors is valuable to provide better understanding as to where future efforts would best be invested.
Our key contributions are as follows:
(a) We provide a detailed analysis of a state-of-the-art pedestrian detector, providing insights into failure cases.
(b) We provide a human baseline for the Caltech Pedestrian Benchmark; as well as a sanitised version of the annotations to serve as new, high quality ground truth for the training and test sets of the benchmark. This data is public111http://www.mpi-inf.mpg.de/pedestrian_detection_cvpr16.
(c) We analyse the effects of training data quality. More specifically we quantify how much better alignment and fewer annotation mistakes can improve performance.
(d) Using the insights of the analysis, we explore variants of top performing methods: filtered channel feature detector  and R-CNN detector [13, 15], and show improvements over the baselines.
Before delving into our analysis, let us describe the datasets in use, their metrics, and our baseline detector.
2.1 Caltech-USA pedestrian detection benchmark
Amongst existing pedestrian datasets [4, 9, 8], KITTI  and Caltech-USA are currently the most popular ones.In this work we focus on the Caltech-USA benchmark  which consists of 2.5 hours of 30Hz video recorded from a vehicle traversing the streets of Los Angeles, USA. The video annotations amount to a total of 350 000 bounding boxes covering unique pedestrians. Detection methods are evaluated on a test set consisting of 4 024 frames. The provided evaluation toolbox generates plots for different subsets of the test set based on annotation size, occlusion level and aspect ratio. The established procedure for training is to use every 30th video frame which results in a total of 4 250 frames with pedestrian cut-outs. More recently, methods which can leverage more data for training have resorted to a finer sampling of the videos [16, 24], yielding up to as much data for training than the standard “” setting.
In the standard Caltech evaluation  the miss rate (MR) is averaged over the low precision range of FPPI (false positives per image). This metric does not reflect well improvements in localisation errors (lowest FPPI range). Aiming for a more complete evaluation, we extend the evaluation FPPI range from traditional to , we denote these and . stands for “original annotations”. In section 3.3 we introduce new annotations, and mark evaluations done there as and . We expect the metric to become more important as detectors get stronger.
2.2 Filtered channel feature detectors
For the analysis in this paper we consider all methods published on the Caltech Pedestrian benchmark, up to the last major conference (CVPR2015). As shown in figure 1, the best method at the time is Checkerboards, and most of the top performing methods are of its same family.
We compare the performance of several detectors from the ICF family in table 2, where we can see a big improvement from 44.2% to 18.5% by introducing filters over the feature channels and optimising the filter bank.
Current top performing convnets methods [15, 20] are sensitive to the underlying detection proposals, thus we first focus on the proposals by optimising the filtered channel feature detectors (more on convnets in section 4.2).
For the experiments involving training new models (in section 4.1) we use our own re-implementation of Checkerboards , based on the LDCF  codebase. To improve the training time we decrease the number of filters from 61 in the original Checkerboards down to 9 filters. Our so-called RotatedFilters are a simplified version of LDCF, applied at three different scales (in the same spirit as SquaresChnFtrs (SCF) ). More details on the filters are given in the supplementary material. As shown in table 2, RotatedFilters are significantly better than the original LDCF, and only (percent point) worse than Checkerboards, yet run faster at training and test time.
3 Analysing the state of the art
In this section we estimate a lower bound on the remaining progress available, analyse the mistakes of current pedestrian detectors, and propose new annotations to better measure future progress.
3.1 Are we reaching saturation?
Progress on pedestrian detection has been showing no sign of slowing in recent years [24, 20, 3], despite recent impressive gains in performance. How much progress can still be expected on current benchmarks? To answer this question, we propose to use a human baseline as lower bound. We asked domain experts to manually “detect” pedestrians in the Caltech-USA test set; machine detection algorithms should be able to at least reach human performance and, eventually, superhuman performance.
Human baseline protocol
To ensure a fair comparison with existing detectors, most of which operate at test time over a single image, we focus on the single frame monocular detection setting. Frames are presented to annotators in random order, and without access to surrounding frames from the source videos. Annotators have to rely on pedestrian appearance and single-frame context rather than (long-term) motion cues.
The Caltech benchmark normalises the aspect ratio of all detection boxes . Thus our human annotations are done by drawing a line from the top of the head to the point between both feet. A bounding box is then automatically generated such that its centre coincides with the centre point of the manually-drawn axis, see illustration in figure 2. This procedure ensures the box is well centred on the subject (which is hard to achieve when marking a bounding box).
To check for consistency among the two annotators, we produced duplicate annotations for a subset of the test images (), and evaluated these separately. With a matching criterion, the results were identical up to a single bounding box.
In figure 3, we compare our human baseline with other top performing methods on different subsets of the test data. We find that the human baseline widely outperforms state-of-the-art detectors in all settings222Except for . This is due to issues with the ground truth, discussed in section 3.3., indicating that there is still room for improvement for automatic methods.
3.2 Failure analysis
Since there is room to grow for existing detectors, one might want to know: when do they fail? In this section we analyse detection mistakes of Checkerboards, which obtains top performance on most subsets of the test set (see figure 3). Since most top methods of figure 1 are of the ICF family, we expect a similar behaviour for them too. Methods using convnets with proposals based on ICF detectors will also be affected.
3.2.1 Error sources
There are two types of errors a detector can do: false positives (detections on background or poorly localised detections) and false negatives (low-scoring or missing pedestrian detections). In this analysis, we look into false positive and false negative detections at 0.1 false positives per image (FPPI, 1 false positive every 10 images), and manually cluster them (one to one mapping) into visually distinctive groups. A total of 402 false positive and 148 false negative detections (missing recall) are categorised by error type.
After inspection, we end up having all false positives clustered in
eleven categories, shown in figure (a)a.
These categories fall into three groups: localisation, background,
and annotation errors. Localisation errors are defined as false detections
overlapping with ground truth bounding boxes, while background errors
have zero overlap with any ground truth annotation.
Background errors are the most common ones, mainly vertical structures (e.g. figure (b)b), tree leaves, and traffic lights. This indicates that the detectors need to be extended with a better vertical context, providing visibility over larger structures and a rough height estimate.
Localisation errors are dominated by double detections (high scoring detections covering the same person, e.g. figure (a)a). This indicates that improved detectors need to have more localised responses (peakier score maps) and/or a different non-maxima suppression strategy. In sections 3.3 and 4.1 we explore how to improve the detector localisation.
The annotation errors are mainly missing ignore regions, and a few missing person annotations. In section 3.3 we revisit the Caltech annotations.
Our clustering results in figure (b)b
show the well known difficulty of detecting small and occluded objects.
We hypothesise that low scoring side-view persons and cyclists may
be due to a dataset bias, i.e. these cases are under-represented in
the training set (most persons are non-cyclist walking on the side-walk,
parallel to the car). Augmenting the training set with external images
for these cases might be an effective strategy.
To understand better the issue with small pedestrians, we measure size, blur, and contrast for each (true or false) detection. We observed that small persons are commonly saturated (over or under exposed) and blurry, and thus hypothesised that this might be an underlying factor for weak detection (other than simply having fewer pixels to make the decision). Our results indicate however that this is not the case. As figure (c)c illustrates, there seems to be no correlation between low detection score and low contrast. This also holds for the blur case, detailed plots are in the supplementary material. We conclude that the small number of pixels is the true source of difficulty. Improving small objects detection thus need to rely on making proper use of all pixels available, both inside the window and in the surrounding context, as well as across time.
Our analysis shows that false positive errors have well defined sources that can be specifically targeted with the strategies suggested above. A fraction of the false negatives are also addressable, albeit the small and occluded pedestrians remain a (hard and) significant problem.
3.2.2 Oracle test cases
The analysis of section 3.2.1 focused on errors counts. For area-under-the-curve metrics, such as the ones used in Caltech, high-scoring errors matter more than low-scoring ones. In this section we directly measure the impact of localisation and background-vs-foreground errors on the detection quality metric (log-average miss-rate) by using oracle test cases.
In the oracle case for localisation, all false positives that overlap with ground truth are ignored for evaluation. In the oracle tests for background-vs-foreground, all false positives that do not overlap with ground truth are ignored.
Figure (a)a shows that fixing localisation mistakes improves performance in the low FPPI region; while fixing background mistakes improves results in the high FPPI region. Fixing both types of mistakes results zero errors, even though this is not immediately visible in the double log plot.
In figure (b)b we show the gains to be obtained in terms by fixing localisation or background issues. When comparing the eight top performing methods we find that most methods would boost performance significantly by fixing either problem. Note that due to the log-log nature of the numbers, the sum of localisation and background deltas do not add up to the total miss-rate.
For most top performing methods localisation and background-vs-foreground errors have equal impact on the detection quality. They are equally important.
3.3 Improved Caltech-USA annotations
When evaluating our human baseline (and other methods) with a strict we notice in figure 3
that the performance drops. The original annotation protocol is based on interpolating sparse annotations across multiple frames, and these sparse annotations are not necessarily located on the evaluated frames. After close inspection we notice that this interpolation generates a systematic offset in the annotations. Humans walk with a natural up and down oscillation that is not modelled by the linear interpolation used, thus in most frames have shifted bounding box annotations. This effect is not noticeable when using the forgiving , however such noise in the annotations is a hurdle when aiming to improve object localisation.
This localisation issues together with the annotation errors detected in section 3.2.1 motivated us to create a new set of improved annotations for the Caltech pedestrians dataset. Our aim is two fold; on one side we want to provide a more accurate evaluation of the state of the art, in particular an evaluation suitable to close the “last 20%” of the problem. On the other side, we want to have training annotations and evaluate how much improved annotations lead to better detections. We evaluate this second aspect in section 4.1.
New annotation protocol
Our new annotations are done both on the test and training
set, and focus on high quality. The annotators are allowed to look
at the full video to decide if a person is present or not, they are
requested to mark ignore regions in areas covering crowds, human shapes
that are not persons (posters, statues, etc.), and in areas that could
not be decided as certainly not containing a person. Each person annotation
is done by drawing a line from the top of the head to the point between
both feet, the same as human baseline. The annotators must hallucinate
head and feet if these are not visible. When the person is not fully
visible, they must also annotate a rectangle around the largest visible
region. This allows to estimate the occlusion level in a similar
fashion as the original annotations. The new annotations do share
some bounding boxes with the human baseline (when no correction was
needed), thus the human baseline cannot be used to do analysis across
different IoU thresholds over the new test set.
In summary, our new annotations differ from the human baseline in the following aspects: both training and test sets are annotated, ignore regions and occlusions are also annotated, full video data is used for decision, and multiple revisions of the same image are allowed.
After creating a full independent set of annotations, we consolidated the new annotations by cross-validating with the old annotations. Any correct old annotation not accounted for in the new set, was added too.
Our new annotations correct several types of errors in the existing annotations, such as misalignments (figure (b)b), missing annotations (false negatives), false annotations (false positives, figure (a)a), and the inconsistent use of “ignore” regions. More examples of “original versus new annotations” provided in the supplementary material, as well as a visualisation software to inspect them frame by frame.
In table 3 we show quantitative evidence that our new annotations are at least more precisely localised than the original ones. We summarise the alignment quality of a detector via the median IoU between true positive detections and a given set of annotations. When evaluating with the original annotations (“median ” column in table 3), only the model trained with original annotations has good localisation. However, when evaluating with the new annotations (“median ” column) both the model trained on INRIA data, and on the new annotations reach high localisation accuracy. This indicates that our new annotations are indeed better aligned, just as INRIA annotations are better aligned than Caltech.
Detailed IoU curves for multiple detectors are provided in the supplementary material. Section 4.1 describes the RotatedFilters-New10 entry.
4 Improving the state of the art
In this section we leverage the insights of the analysis, to improve localisation and background-versus-foreground discrimination of our baseline detector.
4.1 Impact of training annotations
With new annotations at hand we want to understand what is the impact of annotation quality on detection quality. We will train ACF  and RotatedFilters models (introduced in section 2.2) using different training sets and evaluate on both original and new annotations (i.e. , and , ). Note that both detectors are trained via boosting and thus inherently sensitive to annotation noise.
Table 4 shows results when
training with original, new and pruned annotations (using a
training and validation split of the full training set). As expected,
models trained on original/new and tested on original/new perform
better than training and testing on different annotations. To understand
better what the new annotations bring to the table, we build a hybrid
set of annotations. Pruned annotations is a mid-point that allows
to decouple the effects of removing errors and improving alignment.
Pruned annotations are generated by matching new and original annotations (), marking as ignore region any original annotation absent in the new ones, and adding any new annotation absent in the original ones.
From original to pruned annotations the main change is removing annotation errors, from pruned to new, the main change is better alignment. From table 4 both ACF and RotatedFilters benefit from removing annotation errors, even in . This indicates that our new training set is better sanitised than the original one.
We see in that the stronger detector benefits more from better data, and that the largest gain in detection quality comes from removing annotation errors.
The detectors from the ICF family benefit from training with increased
training data [16, 24], using
data is better than (see section 2.1).
To leverage the remaining data using the new
annotations we train a model over the new annotations and use this
model to re-align the original annotations over the portion.
Because the new annotations are better aligned, we expect this model
to be able to recover slight position and scale errors in the original
annotations. Figure 8 shows example results
of this process. See supplementary material for details.
Table 5 reports results using the automatic alignment process, and a few degraded cases: using the original , self-aligning the original using a model trained over original , and aligning the original using only a fraction of the new annotations (without replacing the portion). The results indicate that using a detector model to improve overall data alignment is indeed effective, and that better aligned training data leads to better detection quality (both in and ). This is in line with the analysis of section 3.2. Already using a model trained on of the new annotations for alignment, leads to a stronger model than obtained when using original annotations.
We name the RotatedFilters model trained using the new annotations and the aligned data, RotatedFilters-New10. This model also reaches high median true positives IoU in table 3, indicating that indeed it obtains more precise detections at test time.
|Orig.||Ø||19.20 (34.28)||17.22 (31.65)|
|Orig.||Orig.||19.16 (32.28)||15.71 (28.13)|
|Orig.||New||16.97 (28.01)||14.54 (25.06)|
|New||New||16.77 (29.76)||12.96 (22.20)|
Using high quality annotations for training improves the overall detection quality, thanks both to improved alignment and to reduced annotation errors.
4.2 Convnets for pedestrian detection
The results of section 3.2 indicate that there is room for improvement by focusing on the core background versus foreground discrimination task (the “classification part of object detection”). Recent work [15, 20] showed competitive performance with convolutional neural networks (convnets) for pedestrian detection. We include convnets into our analysis, and explore to what extent performance is driven by the quality of the detection proposals.
AlexNet and VGG
. Both are pre-trained on ImageNet and fine-tuned over Caltech(original annotations) using SquaresChnFtrs proposals. Both networks are based on open source, and both are instances of the R-CNN framework . Albeit their training/test time architectures are slightly different (R-CNN versus Fast R-CNN), we expect the result differences to be dominated by their respective discriminative power (VGG16 improves in mAP over AlexNet in the Pascal detection task ).
|Test proposals||Proposal||+AlexNet||+VGG||+bbox reg|
Table 6 shows that as the quality of the detection proposals improves, AlexNet fails to provide a consistent gain, eventually worsening the results of our ICF detectors (similar observation in ). Similarly VGG provides large gains for weaker proposals, but as the proposals improve, the gain from the convnet re-scoring eventually stalls.
After closer inspection of the resulting curves (see supplementary material), we notice that both AlexNet and VGG push background instances to lower scores, and at the same time generate a large number of high scoring false positives. The ICF detectors are able to provide high recall proposals, where false positives around the objects have low scores (see [15, supp. material, fig. 9]), however convnets have difficulties giving low scores to these windows surrounding the true positives. In other words, despite their fine-tuning, the convnet score maps are “blurrier” than the proposal ones. We hypothesise this is an intrinsic limitation of the AlexNet and VGG architectures, due to their internal feature pooling. Obtaining “peakier” responses from a convnet most likely will require using rather different architectures, possibly more similar to the ones used for semantic labelling or boundaries estimation tasks which require pixel-accurate output.
|RotatedFilters||19.20 (34.28)||17.22 (31.65)|
|+ Alignment §4.1||16.97 (28.01)||14.54 (25.06)|
|+ New annotations §4.1||16.77 (29.76)||12.96 (22.20)|
|+ VGG §4.2||16.61 (34.79)||11.74 (28.37)|
|+ bbox reg & NMS||14.16 (28.39)||10.00 (20.77)|
|Checkerboards||18.47 (33.20)||15.81 (28.57)|
Fortunately, we can compensate for the lack of spatial resolution in the convnet scoring by using bounding box regression. Adding bounding regression over VGG, and applying a second round of non-maximum suppression (first NMS on the proposals, second on the regressed boxes), has the effect of “contracting the score maps”. Neighbour proposals that before generated multiple strong false positives, now collapse into a single high scoring detection. We use the usual merging criterion for the second NMS.
The last column of table 6 shows that bounding box regression + NMS is effective at providing an additional gain over the input proposals, even for our best detector RotatedFilters-New10. On the original annotations RotatedFilters-New10+VGG reaches , which improves over [15, 20]. Our best performing detector RotatedFilters-New10 runs on a image for ~3.5 seconds, including the ICF sliding window detection and VGG rescoring. Training times are counted 1~2 days for the RotatedFilters detector, and 1~2 days for VGG fine-tunning.
Although convnets have strong results in image classification and general object detection, they seem to have limitations when producing well localised detection scores around small objects. Bounding box regression (and NMS) is a key ingredient to side-step this limitation with current architectures. Even after using a strong convnet, background-versus-foreground remains the main source of errors; suggesting that there is still room for improvement on the raw classification power of the neural network.
In this paper, we make great efforts on analysing the failures for a top-performing detector on Caltech dataset. Via our human baseline we have quantified a lower bound on how much improvement there is to be expected. There is a gap still to be closed. To better measure the next steps in detection progress, we have provided new sanitised Caltech train and test set annotations.
Our failure analysis of a top performing method has shown that most of its mistakes are well characterised. The error characteristics lead to specific suggestions on how to engineer better detectors (mentioned in section 3.2; e.g. data augmentation for person side views, or extending the detector receptive field in the vertical axis).
We have partially addressed some of the issues by measuring the impact of better annotations on localisation accuracy, and by investigating the use of convnets to improve the background to foreground discrimination. Our results indicate that significantly better alignment can be achieved with properly trained ICF detectors, and that, for pedestrian detection, convnet struggle with localisation issues, that can be partially addressed via bounding box regression. Both on original and new annotations, the described detection approach reaches top performance, see progress in table 7.
We hope the insights and data provided in this work will guide the path to close the gap between machines and humans in the pedestrian detection task.
-  P. Agrawal, R. Girshick, and J. Malik. Analyzing the performance of multilayer neural networks for object recognition. In ECCV, 2014.
-  R. Benenson, M. Mathias, T. Tuytelaars, and L. Van Gool. Seeking the strongest rigid detector. In CVPR, 2013.
-  R. Benenson, M. Omran, J. Hosang, , and B. Schiele. Ten years of pedestrian detection, what have we learned? In ECCV, CVRSUAD workshop, 2014.
-  N. Dalal and B. Triggs. Histograms of oriented gradients for human detection. In CVPR, 2005.
-  P. Dollár, R. Appel, S. Belongie, and P. Perona. Fast feature pyramids for object detection. PAMI, 2014.
-  P. Dollár, Z. Tu, P. Perona, and S. Belongie. Integral channel features. In BMVC, 2009.
-  P. Dollár, C. Wojek, B. Schiele, and P. Perona. Pedestrian detection: An evaluation of the state of the art. PAMI, 2012.
-  M. Enzweiler and D. M. Gavrila. Monocular pedestrian detection: Survey and experiments. PAMI, 2009.
-  A. Ess, B. Leibe, K. Schindler, and L. Van Gool. A mobile vision system for robust multi-person tracking. In CVPR, 2008.
-  Crete-Roffet F., Dolmiere T., Ladret P., and Nicolas M. The blur effect: Perception and estimation with a new no-reference perceptual blur metric. In SPIE Electronic Imaging Symposium Conf Human Vision and Electronic Imaging, May 2007.
-  A. Geiger, P. Lenz, and R. Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012.
-  R. Girshick. Fast R-CNN. In ICCV, 2015.
-  R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In CVPR, 2014.
-  D. Hoiem, Y. Chodpathumwan, and Q. Dai. Diagnosing error in object detectors. In ECCV, 2012.
-  J. Hosang, M. Omran, R. Benenson, and B. Schiele. Taking a deeper look at pedestrians. In CVPR, 2015.
-  W. Nam, P. Dollár, and J. H. Han. Local decorrelation for improved detection. In NIPS, 2014.
-  W. Ouyang and X. Wang. Single-pedestrian detection aided by multi-pedestrian detection. In CVPR, 2013.
-  S. Paisitkriangkrai, C. Shen, and A. van den Hengel. Strengthening the effectiveness of pedestrian detection with spatially pooled features. In ECCV, 2014.
D. Park, C. L. Zitnick, D. Ramanan, and P. Dollár.
Exploring weak stabilization for motion feature extraction.In CVPR, 2013.
Y. Tian, P. Luo, X. Wang, and X. Tang.
Pedestrian detection aided by deep learning semantic tasks.In CVPR, 2015.
-  C. Wojek and B. Schiele. A performance evaluation of single and multi-feature people detection. In DAGM, May 2008.
-  S. Zhang, C. Bauckhage, and A. B. Cremers. Informed haar-like features improve pedestrian detection. In CVPR, 2014.
-  S. Zhang, C. Bauckhage, D. A. Klein, and A. B. Cremers. Exploring human vision driven features for pedestrian detection. IEEE Transactions on Circuits and Systems for Video Technology, 2015.
-  S. Zhang, R. Benenson, and B. Schiele. Filtered channel features for pedestrian detection. In CVPR, 2015.
Appendix A Content
This supplementary material provides a more detailed view of some of the aspects presented in the main paper.
Section B gives details of the RotatedFilters detector we used for our experiments (section 2.2 in main paper).
Section C provides the detailed curves behind the summary bar plots for different test set subsets (see figure 3 and section 3.1 in main paper).
Section D shows examples for each error type from the analysed detector, discusses the scale, blur and contrast evaluations, and revisits the oracle cases experiments in more detail (section 3.2 in main paper).
Section E shows examples of how the new training annotations improve over the original ones (section 3.3 in main paper).
Section F discuss the impact of new annotations on the evaluation of existing methods (MR ranking and recall-versus-IoU curves) (section 4.1 in main paper).
Section G shows the effects of automatically aligning data with data (section 4.1 in main paper).
Figure 26 summarises our final detection results both in original and new annotations.
Appendix B Rotated filters detector
For our experiments we re-implement the filtered channel feature Checkerboards detector  using the LDCF  codebase. The training procedure turns out to be slow due to the large number of filters (61 filters per channel). To accelerate the training and test procedures, we design a small set of 9 filters per channel that still provides good performance. We call our new filtered channel feature detector; RotatedFilters (see figure (d)d).
The rotated filters are inspired by the filterbank of LDCF (obtained by applying PCA to each feature channel). The first three filters of LDCF of each features channel are the constant filter and two step functions in orthogonal directions, with the particularities that the oriented gradient channels also have rotated filters (see figure (b)b). Our rotated filters are stylised versions of LDCF. The resulting RotatedFilters filterbank is somewhat intuitive, while filters from Checkerboards, are less systematic and less clear in their function (see figure (c)c).
On the Caltech validation set, RotatedFilters obtains 31.6% using one scale (4x4); and 28.9% using three scales (4x4, 8x8 and 16x16). Therefore, we select this 3-scale structure in our experiments. On the test set, the performance of RotatedFilters is 19.2% , i.e. a less than 1% loss with respect to Checkerboards, yet it is ~6x faster at feature computation.
In this paper, we use RotatedFilters for all experiments involving training a new model.
Appendix C Results per test subset
Figure 12 contains the detailed curves behind figure 3 in the main paper (“subsets bar plot”). We can see that Checkerboards and RotatedFilters show good performance across all subsets. The few cases where they are not top ranked (e.g. figures (e)e and (h)h) all methods exhibit low detection quality, and thus all have similarly poor scores.
Figure 12 shows that Checkerboards is not optimised for the most common case on the Caltech dataset, but instead shows good performance across a variety of situations; and is thus an interesting method to analyse.
Appendix D Checkerboards errors analysis
Blur and contrast measures
To enable our analysis regarding blur and contrast, we define two automated measures. We measure blur using the method from 
, while contrast is measured via the difference between the top and bottom quantiles of the grey scale intensity of the pedestrian patch.
Figures 15 and 16 show pedestrians ranked by our blur and contrast measures. One can observe that our quantitative measures correlate well with the qualitative notions of blur and contrast.
Scale, blur, or contrast?
For false negatives, a major source of error is small scale, but we find small pedestrians are often of low contrast or blurred. In order to investigate the three factors separately, we observe the correlation between size/contrast/blur and score, as shown in figure 14. We can see that the overlap between false positive and true positive is equally distributed across different levels of contrast and blur; while for scale, the overlap is quite dense at small scale. To this end, we conclude that small scale is the main factor negatively impacting detection quality; and that blur and contrast are uninformative measures for the detection task.
d.1 Oracle cases
In figure 6, we show the standard evaluation and oracle evaluation curves for state-of-the-art methods. For the localisation oracle, false positives that overlap with the ground truth are not considered; for the background-versus-foreground oracle, false positives that do not overlap with the ground truth are not considered. Based on the curves, we have the following findings:
All methods are significantly improved in each oracle evaluation.
The ranking of all methods stays relatively stable in each oracle case.
In terms of MR, the improvement is comparable for localisation or background-versus-foreground oracle tests; the detection performance can be boosted by fixing either problem.
We also show some examples of objects with similar scores in figure 13. In both low-scoring and high-scoring groups, we can see both pedestrians and background objects, which shows that the detector fails to rank foreground and background adequately.
d.2 Log scale visual distortion
In the paper we show results for so called oracle experiments that emulate the case in which we do not make one type of error: we remove either mistakes that touch annotated pedestrians (localisation oracle) or mistakes that are located on background (background oracle).
It is important to note that these are the only two types of false positives. If we remove both types the only mistakes that remain stem from missing recall and the result would be a horizontal line with very low miss rate.
Because of the double log scale in the performance plots on Caltech the curves look like both oracles improve performance slightly but the bulk of mistakes arise from a different type of mistakes, which is not the case.
In figure 22 we illustrate how much double log scales distort areas. We often think of the average miss rate as the area under the curve, so we colour code the false positives in the plots by their type: the plot shows the ratio between localisation (blue) and background (green) mistakes at every point on the miss rate, but also for the entire curve. Both curves, (b)b and (c)c show the same data with the only difference that one shows localisations on the left and the other one on the right. Due to the double log scale, the error type that is plotted on the left seems to dominate the metric.
Appendix E Improved annotations
In figure 23 we show original (red) and new annotations (green) on example frames from the test set. From the comparison, we can see that the new annotations are better aligned to the pedestrians. This results from the fact that head and feet are closer to the centre of the new bounding boxes.
Appendix F Evaluation on original and new annotations
Figure 25 presents the ranking of all published Caltech methods up to CVPR 2015 when evaluated on (original annotations), or on (proposed new annotations). Although there are a few changes in ranking (e.g. JointDeep versus SDN), the overall trend is preserved. This is a good sign that the improved annotations are not a radical departure from previous ones. As discussed in the paper (and in other sections of the supplementary material), improved annotations matter most for future methods (going further down in MR), and for the low FPPI region of the curves (high confidence mistakes).
MR versus IoU
Section 3.3 (and table 3) of the main paper discuss an empirical measure
of how the new annotations are better aligned. Here we provide some
Figure 24 plots and of top performing methods versus the overlap criterion for accepting detections as true positives (IoU threshold). The standard evaluation uses IoU threshold . On these plots methods trained on INRIA have continuous lines, methods trained on Caltech dashed ones (see also figure 25).
In figure (a)a (original annotations) the ranking of the methods remains stable as the overlap threshold becomes stricter (consistent with the observations in ). Interestingly we observe a different trend in figure (b)b (new annotations).
When evaluating (new annotations) we see that methods training on INRIA, albeit having a poor performance at , perform comparatively well at higher IoU, eventually overpassing all methods trained on raw Caltech data. We attribute this to the fact that INRIA training data is of better quality (better aligned training samples), and thus the detectors have learnt to localise better. This difference in trend between original and new annotations confirms that our improved annotations are better with respect to localization. Table 3 in the main paper provides a summarised version of figure 24.
Appendix G Impact of aligning Caltech
We can see from (b)b that using our semi-automatically aligned Caltech training data provides a significant boost in localization quality. From RotatedFilters to RotatedFilters-New10x the improves across the full IoU range. Figure 27 shows qualitative results for the alignment procedure done over the training data.