Deep Convolutional Neural Networks (DCNNs) have recently shown state of the art performance in high level vision tasks, such as image classification and object detection. This work brings together methods from DCNNs and probabilistic graphical models for addressing the task of pixel-level classification (also called "semantic image segmentation"). We show that responses at the final layer of DCNNs are not sufficiently localized for accurate object segmentation. This is due to the very invariance properties that make DCNNs good for high level tasks. We overcome this poor localization property of deep networks by combining the responses at the final DCNN layer with a fully connected Conditional Random Field (CRF). Qualitatively, our "DeepLab" system is able to localize segment boundaries at a level of accuracy which is beyond previous methods. Quantitatively, our method sets the new state-of-art at the PASCAL VOC-2012 semantic image segmentation task, reaching 71.6 efficiently: Careful network re-purposing and a novel application of the 'hole' algorithm from the wavelet community allow dense computation of neural net responses at 8 frames per second on a modern GPU.READ FULL TEXT VIEW PDF
In this work we address the task of semantic image segmentation with Dee...
Convolutional neural networks with many layers have recently been shown ...
Deep convolutional neural networks (CNNs) are the backbone of state-of-a...
There has been significant interest in the use of fully-connected graphi...
While significant attention has been recently focused on designing super...
This paper addresses semantic image segmentation by incorporating rich
Supervised training of a convolutional network for object classification...
Working copy of Deeplab repository
Generic Foreground Segmentation in Images
Deep-Lab combines CNN with CRF. It's built on top of caffe
Holed Convolution Layer for Semantic Segmentation in MatConvNet
Deep Convolutional Neural Networks (DCNNs) had been the method of choice for document recognition since LeCun et al. (1998)
, but have only recently become the mainstream of high-level vision research. Over the past two years DCNNs have pushed the performance of computer vision systems to soaring heights on a broad array of high-level problems, including image classification(Krizhevsky et al., 2013; Sermanet et al., 2013; Simonyan & Zisserman, 2014; Szegedy et al., 2014; Papandreou et al., 2014), object detection (Girshick et al., 2014), fine-grained categorization (Zhang et al., 2014), among others. A common theme in these works is that DCNNs trained in an end-to-end manner deliver strikingly better results than systems relying on carefully engineered representations, such as SIFT or HOG features. This success can be partially attributed to the built-in invariance of DCNNs to local image transformations, which underpins their ability to learn hierarchical abstractions of data (Zeiler & Fergus, 2014)
. While this invariance is clearly desirable for high-level vision tasks, it can hamper low-level tasks, such as pose estimation(Chen & Yuille, 2014; Tompson et al., 2014) and semantic segmentation - where we want precise localization, rather than abstraction of spatial details.
There are two technical hurdles in the application of DCNNs to image labeling tasks: signal downsampling, and spatial ‘insensitivity’ (invariance). The first problem relates to the reduction of signal resolution incurred by the repeated combination of max-pooling and downsampling (‘striding’) performed at every layer of standard DCNNs(Krizhevsky et al., 2013; Simonyan & Zisserman, 2014; Szegedy et al., 2014). Instead, as in Papandreou et al. (2014), we employ the ‘atrous’ (with holes) algorithm originally developed for efficiently computing the undecimated discrete wavelet transform (Mallat, 1999). This allows efficient dense computation of DCNN responses in a scheme substantially simpler than earlier solutions to this problem (Giusti et al., 2013; Sermanet et al., 2013).
The second problem relates to the fact that obtaining object-centric decisions from a classifier requires invariance to spatial transformations, inherently limiting the spatial accuracy of the DCNN model. We boost our model’s ability to capture fine details by employing a fully-connected Conditional Random Field (CRF). Conditional Random Fields have been broadly used in semantic segmentation to combine class scores computed by multi-way classifiers with the low-level information captured by the local interactions of pixels and edges(Rother et al., 2004; Shotton et al., 2009) or superpixels (Lucchi et al., 2011). Even though works of increased sophistication have been proposed to model the hierarchical dependency (He et al., 2004; Ladicky et al., 2009; Lempitsky et al., 2011) and/or high-order dependencies of segments (Delong et al., 2012; Gonfaus et al., 2010; Kohli et al., 2009; Chen et al., 2013; Wang et al., 2015), we use the fully connected pairwise CRF proposed by Krähenbühl & Koltun (2011) for its efficient computation, and ability to capture fine edge details while also catering for long range dependencies. That model was shown in Krähenbühl & Koltun (2011) to largely improve the performance of a boosting-based pixel-level classifier, and in our work we demonstrate that it leads to state-of-the-art results when coupled with a DCNN-based pixel-level classifier.
The three main advantages of our “DeepLab” system are (i) speed: by virtue of the ‘atrous’ algorithm, our dense DCNN operates at 8 fps, while Mean Field Inference for the fully-connected CRF requires 0.5 second, (ii) accuracy: we obtain state-of-the-art results on the PASCAL semantic segmentation challenge, outperforming the second-best approach of Mostajabi et al. (2014) by a margin of 7.2 and (iii) simplicity: our system is composed of a cascade of two fairly well-established modules, DCNNs and CRFs.
Our system works directly on the pixel representation, similarly to Long et al. (2014). This is in contrast to the two-stage approaches that are now most common in semantic segmentation with DCNNs: such techniques typically use a cascade of bottom-up image segmentation and DCNN-based region classification, which makes the system commit to potential errors of the front-end segmentation system. For instance, the bounding box proposals and masked regions delivered by (Arbeláez et al., 2014; Uijlings et al., 2013) are used in Girshick et al. (2014) and (Hariharan et al., 2014b) as inputs to a DCNN to introduce shape information into the classification process. Similarly, the authors of Mostajabi et al. (2014) rely on a superpixel representation. A celebrated non-DCNN precursor to these works is the second order pooling method of (Carreira et al., 2012) which also assigns labels to the regions proposals delivered by (Carreira & Sminchisescu, 2012). Understanding the perils of committing to a single segmentation, the authors of Cogswell et al. (2014) build on (Yadollahpour et al., 2013) to explore a diverse set of CRF-based segmentation proposals, computed also by (Carreira & Sminchisescu, 2012). These segmentation proposals are then re-ranked according to a DCNN trained in particular for this reranking task. Even though this approach explicitly tries to handle the temperamental nature of a front-end segmentation algorithm, there is still no explicit exploitation of the DCNN scores in the CRF-based segmentation algorithm: the DCNN is only applied post-hoc, while it would make sense to directly try to use its results during segmentation.
Moving towards works that lie closer to our approach, several other researchers have considered the use of convolutionally computed DCNN features for dense image labeling. Among the first have been Farabet et al. (2013) who apply DCNNs at multiple image resolutions and then employ a segmentation tree to smooth the prediction results; more recently, Hariharan et al. (2014a) propose to concatenate the computed inter-mediate feature maps within the DCNNs for pixel classification, and Dai et al. (2014) propose to pool the inter-mediate feature maps by region proposals. Even though these works still employ segmentation algorithms that are decoupled from the DCNN classifier’s results, we believe it is advantageous that segmentation is only used at a later stage, avoiding the commitment to premature decisions.
More recently, the segmentation-free techniques of (Long et al., 2014; Eigen & Fergus, 2014) directly apply DCNNs to the whole image in a sliding window fashion, replacing the last fully connected layers of a DCNN by convolutional layers. In order to deal with the spatial localization issues outlined in the beginning of the introduction, Long et al. (2014) upsample and concatenate the scores from inter-mediate feature maps, while Eigen & Fergus (2014) refine the prediction result from coarse to fine by propagating the coarse results to another DCNN.
The main difference between our model and other state-of-the-art models is the combination of pixel-level CRFs and DCNN-based ‘unary terms’. Focusing on the closest works in this direction, Cogswell et al. (2014) use CRFs as a proposal mechanism for a DCNN-based reranking system, while Farabet et al. (2013) treat superpixels as nodes for a local pairwise CRF and use graph-cuts for discrete inference; as such their results can be limited by errors in superpixel computations, while ignoring long-range superpixel dependencies. Our approach instead treats every pixel as a CRF node, exploits long-range dependencies, and uses CRF inference to directly optimize a DCNN-driven cost function. We note that mean field had been extensively studied for traditional image segmentation/edge detection tasks, e.g., (Geiger & Girosi, 1991; Geiger & Yuille, 1991; Kokkinos et al., 2008), but recently Krähenbühl & Koltun (2011) showed that the inference can be very efficient for fully connected CRF and particularly effective in the context of semantic segmentation.
After the first version of our manuscript was made publicly available, it came to our attention that two other groups have independently and concurrently pursued a very similar direction, combining DCNNs and densely connected CRFs (Bell et al., 2014; Zheng et al., 2015). There are several differences in technical aspects of the respective models. Bell et al. (2014) focus on the problem of material classification, while Zheng et al. (2015) unroll the CRF mean-field inference steps to convert the whole system into an end-to-end trainable feed-forward network.
We have updated our proposed “DeepLab” system with much improved methods and results in our latest work (Chen et al., 2016). We refer the interested reader to the paper for details.
Herein we describe how we have re-purposed and finetuned the publicly available Imagenet-pretrained state-of-art 16-layer classification network of(Simonyan & Zisserman, 2014) (VGG-16) into an efficient and effective dense feature extractor for our dense semantic image segmentation system.
Dense spatial score evaluation is instrumental in the success of our dense CNN feature extractor. As a first step to implement this, we convert the fully-connected layers of VGG-16 into convolutional ones and run the network in a convolutional fashion on the image at its original resolution. However this is not enough as it yields very sparsely computed detection scores (with a stride of 32 pixels). To compute scores more densely at our target stride of 8 pixels, we develop a variation of the method previously employed by Giusti et al. (2013); Sermanet et al. (2013). We skip subsampling after the last two max-pooling layers in the network of Simonyan & Zisserman (2014) and modify the convolutional filters in the layers that follow them by introducing zeros to increase their length ( in the last three convolutional layers and in the first fully connected layer). We can implement this more efficiently by keeping the filters intact and instead sparsely sample the feature maps on which they are applied on using an input stride of 2 or 4 pixels, respectively. This approach, illustrated in Fig. 1 is known as the ‘hole algorithm’ (‘atrous algorithm’) and has been developed before for efficient computation of the undecimated wavelet transform (Mallat, 1999)
. We have implemented this within the Caffe framework(Jia et al., 2014) by adding to the im2col
function (it converts multi-channel feature maps to vectorized patches) the option to sparsely sample the underlying feature map. This approach is generally applicable and allows us to efficiently compute dense CNN feature maps at any target subsampling rate without introducing any approximations.
We finetune the model weights of the Imagenet-pretrained VGG-16 network to adapt it to the image classification task in a straightforward fashion, following the procedure of Long et al. (2014)
. We replace the 1000-way Imagenet classifier in the last layer of VGG-16 with a 21-way one. Our loss function is the sum of cross-entropy terms for each spatial position in the CNN output map (subsampled by 8 compared to the original image). All positions and labels are equally weighted in the overall loss function. Our targets are the ground truth labels (subsampled by 8). We optimize the objective function with respect to the weights at all network layers by the standard SGD procedure ofKrizhevsky et al. (2013).
, the class score maps (corresponding to log-probabilities) are quite smooth, which allows us to use simple bilinear interpolation to increase their resolution by a factor of 8 at a negligible computational cost. Note that the method ofLong et al. (2014) does not use the hole algorithm and produces very coarse scores (subsampled by a factor of 32) at the CNN output. This forced them to use learned upsampling layers, significantly increasing the complexity and training time of their system: Fine-tuning our network on PASCAL VOC 2012 takes about 10 hours, while they report a training time of several days (both timings on a modern GPU).
Another key ingredient in re-purposing our network for dense score computation is explicitly controlling the network’s receptive field size. Most recent DCNN-based image recognition methods rely on networks pre-trained on the Imagenet large-scale classification task. These networks typically have large receptive field size: in the case of the VGG-16 net we consider, its receptive field is
(with zero-padding) andpixels if the net is applied convolutionally. After converting the network to a fully convolutional one, the first fully connected layer has 4,096 filters of large spatial size and becomes the computational bottleneck in our dense score map computation.
We have addressed this practical problem by spatially subsampling (by simple decimation) the first FC layer to (or ) spatial size. This has reduced the receptive field of the network down to (with zero-padding) or (in convolutional mode) and has reduced computation time for the first FC layer by times. Using our Caffe-based implementation and a Titan GPU, the resulting VGG-derived network is very efficient: Given a input image, it produces dense raw feature scores at the top of the network at a rate of about 8 frames/sec during testing. The speed during training is 3 frames/sec. We have also successfully experimented with reducing the number of channels at the fully connected layers from 4,096 down to 1,024, considerably further decreasing computation time and memory footprint without sacrificing performance, as detailed in Section 5. Using smaller networks such as Krizhevsky et al. (2013) could allow video-rate test-time dense feature computation even on light-weight GPUs.
As illustrated in Figure 2, DCNN score maps can reliably predict the presence and rough position of objects in an image but are less well suited for pin-pointing their exact outline. There is a natural trade-off between classification accuracy and localization accuracy with convolutional networks: Deeper models with multiple max-pooling layers have proven most successful in classification tasks, however their increased invariance and large receptive fields make the problem of inferring position from the scores at their top output levels more challenging.
Recent work has pursued two directions to address this localization challenge. The first approach is to harness information from multiple layers in the convolutional network in order to better estimate the object boundaries (Long et al., 2014; Eigen & Fergus, 2014). The second approach is to employ a super-pixel representation, essentially delegating the localization task to a low-level segmentation method. This route is followed by the very successful recent method of Mostajabi et al. (2014).
In Section 4.2, we pursue a novel alternative direction based on coupling the recognition capacity of DCNNs and the fine-grained localization accuracy of fully connected CRFs and show that it is remarkably successful in addressing the localization challenge, producing accurate semantic segmentation results and recovering object boundaries at a level of detail that is well beyond the reach of existing methods.
|Image/G.T.||DCNN output||CRF Iteration 1||CRF Iteration 2||CRF Iteration 10|
Traditionally, conditional random fields (CRFs) have been employed to smooth noisy segmentation maps (Rother et al., 2004; Kohli et al., 2009). Typically these models contain energy terms that couple neighboring nodes, favoring same-label assignments to spatially proximal pixels. Qualitatively, the primary function of these short-range CRFs has been to clean up the spurious predictions of weak classifiers built on top of local hand-engineered features.
Compared to these weaker classifiers, modern DCNN architectures such as the one we use in this work produce score maps and semantic label predictions which are qualitatively different. As illustrated in Figure 2, the score maps are typically quite smooth and produce homogeneous classification results. In this regime, using short-range CRFs can be detrimental, as our goal should be to recover detailed local structure rather than further smooth it. Using contrast-sensitive potentials (Rother et al., 2004) in conjunction to local-range CRFs can potentially improve localization but still miss thin-structures and typically requires solving an expensive discrete optimization problem.
To overcome these limitations of short-range CRFs, we integrate into our system the fully connected CRF model of Krähenbühl & Koltun (2011). The model employs the energy function
where is the label assignment for pixels. We use as unary potential , where is the label assignment probability at pixel as computed by DCNN. The pairwise potential is , where , and zero otherwise (i.e., Potts Model). There is one pairwise term for each pair of pixels and in the image no matter how far from each other they lie, i.e. the model’s factor graph is fully connected. Each is the Gaussian kernel depends on features (denoted as ) extracted for pixel and and is weighted by parameter . We adopt bilateral position and color terms, specifically, the kernels are
where the first kernel depends on both pixel positions (denoted as ) and pixel color intensities (denoted as ), and the second kernel only depends on pixel positions. The hyper parameters , and control the “scale” of the Gaussian kernels.
Crucially, this model is amenable to efficient approximate probabilistic inference (Krähenbühl & Koltun, 2011). The message passing updates under a fully decomposable mean field approximation can be expressed as convolutions with a Gaussian kernel in feature space. High-dimensional filtering algorithms (Adams et al., 2010) significantly speed-up this computation resulting in an algorithm that is very fast in practice, less that 0.5 sec on average for Pascal VOC images using the publicly available implementation of (Krähenbühl & Koltun, 2011).
we have also explored a multi-scale prediction method to increase the boundary localization accuracy. Specifically, we attach to the input image and the output of each of the first four max pooling layers a two-layer MLP (first layer: 128 3x3 convolutional filters, second layer: 128 1x1 convolutional filters) whose feature map is concatenated to the main network’s last layer feature map. The aggregate feature map fed into the softmax layer is thus enhanced by 5 * 128 = 640 channels. We only adjust the newly added weights, keeping the other network parameters to the values learned by the method of Section3. As discussed in the experimental section, introducing these extra direct connections from fine-resolution layers improves localization performance, yet the effect is not as dramatic as the one obtained with the fully-connected CRF.
We test our DeepLab model on the PASCAL VOC 2012 segmentation benchmark (Everingham et al., 2014), consisting of 20 foreground object classes and one background class. The original dataset contains , , and images for training, validation, and testing, respectively. The dataset is augmented by the extra annotations provided by Hariharan et al. (2011), resulting in training images. The performance is measured in terms of pixel intersection-over-union (IOU) averaged across the 21 classes.
We adopt the simplest form of piecewise training, decoupling the DCNN and CRF training stages, assuming the unary terms provided by the DCNN are fixed during CRF training.
For DCNN training we employ the VGG-16 network which has been pre-trained on ImageNet. We fine-tuned the VGG-16 network on the VOC 21-way pixel-classification task by stochastic gradient descent on the cross-entropy loss function, as described in Section3.1. We use a mini-batch of 20 images and initial learning rate of ( for the final classifier layer), multiplying the learning rate by 0.1 at every 2000 iterations. We use momentum of and a weight decay of .
After the DCNN has been fine-tuned, we cross-validate the parameters of the fully connected CRF model in Eq. (2) along the lines of Krähenbühl & Koltun (2011). We use the default values of and and we search for the best values of , , and by cross-validation on a small subset of the validation set (we use 100 images). We employ coarse-to-fine search scheme. Specifically, the initial search range of the parameters are , and (MATLAB notation), and then we refine the search step sizes around the first round’s best values. We fix the number of mean field iterations to 10 for all reported experiments.
|Method mean IOU (%) DeepLab 59.80 DeepLab-CRF 63.74 DeepLab-MSc 61.30 DeepLab-MSc-CRF 65.21 DeepLab-7x7 64.38 DeepLab-CRF-7x7 67.64 DeepLab-LargeFOV 62.25 DeepLab-CRF-LargeFOV 67.64 DeepLab-MSc-LargeFOV 64.21 DeepLab-MSc-CRF-LargeFOV 68.70||Method mean IOU (%) MSRA-CFM 61.8 FCN-8s 62.2 TTI-Zoomout-16 64.4 DeepLab-CRF 66.4 DeepLab-MSc-CRF 67.1 DeepLab-CRF-7x7 70.3 DeepLab-CRF-LargeFOV 70.3 DeepLab-MSc-CRF-LargeFOV 71.6|
We conduct the majority of our evaluations on the PASCAL ‘val’ set, training our model on the augmented PASCAL ‘train’ set. As shown in Tab. 1 (a), incorporating the fully connected CRF to our model (denoted by DeepLab-CRF) yields a substantial performance boost, about 4% improvement over DeepLab. We note that the work of Krähenbühl & Koltun (2011) improved the result of TextonBoost (Shotton et al., 2009) to , which makes the improvement we report here (from to ) all the more impressive.
Turning to qualitative results, we provide visual comparisons between DeepLab and DeepLab-CRF in Fig. 7. Employing a fully connected CRF significantly improves the results, allowing the model to accurately capture intricate object boundaries.
We also exploit the features from the intermediate layers, similar to Hariharan et al. (2014a); Long et al. (2014). As shown in Tab. 1 (a), adding the multi-scale features to our DeepLab model (denoted as DeepLab-MSc) improves about performance, and further incorporating the fully connected CRF (denoted as DeepLab-MSc-CRF) yields about 4% improvement. The qualitative comparisons between DeepLab and DeepLab-MSc are shown in Fig. 4. Leveraging the multi-scale features can slightly refine the object boundaries.
The ‘atrous algorithm’ we employed allows us to arbitrarily control the Field-of-View (FOV) of the models by adjusting the input stride, as illustrated in Fig. 1. In Tab. 2, we experiment with several kernel sizes and input strides at the first fully connected layer. The method, DeepLab-CRF-7x7, is the direct modification from VGG-16 net, where the kernel size = and input stride = 4. This model yields performance of on the ‘val’ set, but it is relatively slow ( images per second during training). We have improved model speed to images per second by reducing the kernel size to . We have experimented with two such network variants with different FOV sizes, DeepLab-CRF and DeepLab-CRF-4x4; the latter has large FOV (i.e., large input stride) and attains better performance. Finally, we employ kernel size and input stride = 12, and further change the filter sizes from 4096 to 1024 for the last two layers. Interestingly, the resulting model, DeepLab-CRF-LargeFOV, matches the performance of the expensive DeepLab-CRF-7x7. At the same time, it is times faster to run and has significantly fewer parameters (20.5M instead of 134.3M).
The performance of several model variants is summarized in Tab. 1, showing the benefit of exploiting multi-scale features and large FOV.
|Method||kernel size||input stride||receptive field||# parameters||mean IOU (%)||Training speed (img/sec)|
To quantify the accuracy of the proposed model near object boundaries, we evaluate the segmentation accuracy with an experiment similar to Kohli et al. (2009); Krähenbühl & Koltun (2011). Specifically, we use the ‘void’ label annotated in val set, which usually occurs around object boundaries. We compute the mean IOU for those pixels that are located within a narrow band (called trimap) of ‘void’ labels. As shown in Fig. 5, exploiting the multi-scale features from the intermediate layers and refining the segmentation results by a fully connected CRF significantly improve the results around object boundaries.
We have implemented the proposed methods by extending the excellent Caffe framework (Jia et al., 2014). We share our source code, configuration files, and trained models that allow reproducing the results in this paper at a companion web site https://bitbucket.org/deeplab/deeplab-public.
|(a) FCN-8s vs. DeepLab-CRF||(b) TTI-Zoomout-16 vs. DeepLab-CRF|
Having set our model choices on the validation set, we evaluate our model variants on the PASCAL VOC 2012 official ‘test’ set. As shown in Tab. 3, our DeepLab-CRF and DeepLab-MSc-CRF models achieve performance of and mean IOU111http://host.robots.ox.ac.uk:8080/leaderboard/displaylb.php?challengeid=11&compid=6, respectively. Our models outperform all the other state-of-the-art models (specifically, TTI-Zoomout-16 (Mostajabi et al., 2014), FCN-8s (Long et al., 2014), and MSRA-CFM (Dai et al., 2014)). When we increase the FOV of the models, DeepLab-CRF-LargeFOV yields performance of , the same as DeepLab-CRF-7x7, while its training speed is faster. Furthermore, our best model, DeepLab-MSc-CRF-LargeFOV, attains the best performance of by employing both multi-scale features and large FOV.
Our work combines ideas from deep convolutional neural networks and fully-connected conditional random fields, yielding a novel method able to produce semantically accurate predictions and detailed segmentation maps, while being computationally efficient. Our experimental results show that the proposed method significantly advances the state-of-art in the challenging PASCAL VOC 2012 semantic image segmentation task.
There are multiple aspects in our model that we intend to refine, such as fully integrating its two main components (CNN and CRF) and train the whole system in an end-to-end fashion, similar to Krähenbühl & Koltun (2013); Chen et al. (2014); Zheng et al. (2015). We also plan to experiment with more datasets and apply our method to other sources of data such as depth maps or videos. Recently, we have pursued model training with weakly supervised annotations, in the form of bounding boxes or image-level labels (Papandreou et al., 2015).
At a higher level, our work lies in the intersection of convolutional neural networks and probabilistic graphical models. We plan to further investigate the interplay of these two powerful classes of methods and explore their synergistic potential for solving challenging computer vision tasks.
This work was partly supported by ARO 62250-CS, NIH Grant 5R01EY022247-03, EU Project RECONFIG FP7-ICT-600825 and EU Project MOBOT FP7-ICT-2011-600796. We also gratefully acknowledge the support of NVIDIA Corporation with the donation of GPUs used for this research. We would like to thank the anonymous reviewers for their detailed comments and constructive feedback.
Here we present the list of major paper revisions for the convenience of the readers.
Submission to ICLR 2015. Introduces the model DeepLab-CRF, which attains the performance of on PASCAL VOC 2012 test set.
Rebuttal for ICLR 2015. Adds the model DeepLab-MSc-CRF, which incorporates multi-scale features from the intermediate layers. DeepLab-MSc-CRF yields the performance of on PASCAL VOC 2012 test set.
Camera-ready for ICLR 2015. Experiments with large Field-Of-View. On PASCAL VOC 2012 test set, DeepLab-CRF-LargeFOV achieves the performance of . When exploiting both mutli-scale features and large FOV, DeepLab-MSc-CRF-LargeFOV attains the performance of .
Reference to our updated “DeepLab” system (Chen et al., 2016) with much improved results.