{"title": "Grid Saliency for Context Explanations of Semantic Segmentation", "book": "Advances in Neural Information Processing Systems", "page_first": 6462, "page_last": 6473, "abstract": "Recently, there has been a growing interest in developing saliency methods that provide visual explanations of network predictions. Still, the usability of existing methods is limited to image classification models. To overcome this limitation, we extend the existing approaches to generate grid saliencies, which provide spatially coherent visual explanations for (pixel-level) dense prediction networks. As the proposed grid saliency allows to spatially disentangle the object and its context, we specifically explore its potential to produce context explanations for semantic segmentation networks, discovering which context most influences the class predictions inside a target object area. We investigate the effectiveness of grid saliency on a synthetic dataset with an artificially induced bias between objects and their context as well as on the real-world Cityscapes dataset using state-of-the-art segmentation networks. Our results show that grid saliency can be successfully used to provide easily interpretable context explanations and, moreover, can be employed for detecting and localizing contextual biases present in the data.", "full_text": "Grid Saliency for\n\nContext Explanations of Semantic Segmentation\n\nLukas Hoyer\n\nMauricio Munoz\n\nPrateek Katiyar\n\nAnna Khoreva\nVolker Fischer\nBosch Center for Arti\ufb01cial Intelligence\n\nlukas.hoyer@outlook.com\n\n{firstname.lastname}@bosch.com\n\nAbstract\n\nRecently, there has been a growing interest in developing saliency methods that\nprovide visual explanations of network predictions. Still, the usability of existing\nmethods is limited to image classi\ufb01cation models. To overcome this limitation,\nwe extend the existing approaches to generate grid saliencies, which provide\nspatially coherent visual explanations for (pixel-level) dense prediction networks.\nAs the proposed grid saliency allows to spatially disentangle the object and its\ncontext, we speci\ufb01cally explore its potential to produce context explanations for\nsemantic segmentation networks, discovering which context most in\ufb02uences the\nclass predictions inside a target object area. We investigate the effectiveness of grid\nsaliency on a synthetic dataset with an arti\ufb01cially induced bias between objects and\ntheir context as well as on the real-world Cityscapes dataset using state-of-the-art\nsegmentation networks. Our results show that grid saliency can be successfully\nused to provide easily interpretable context explanations and, moreover, can be\nemployed for detecting and localizing contextual biases present in the data.\n\n1\n\nIntroduction\n\nIn many real-world scenarios, the presence of an object, its location and appearance are highly\ncorrelated with the contextual information surrounding this object, such as the presence of other\nnearby objects or more global scene semantics. For example, in the case of an urban street scene,\na cyclist is more likely to co-occur on a bicycle and a car to appear on the road below sky and\nbuildings (cf. objects and their context explanations in Fig. 1). These semantic correlations are\ninherently present in real-world data. A data-driven model, such as a deep neural network, is prone\nto exploit these statistical biases during training in order to improve its prediction performance.\nThese biases picked up by the network during training can lead to erroneous predictions and impair\nnetwork generalization (cf. the effect of context on misclassi\ufb01cation in Fig. 2). An effective and safe\nutilization of deep learning models for real-world applications, e.g. autonomous driving, requires a\ngood understanding of these contextual biases inherent in the data and the extent to which a learned\nmodel incorporated them into its decision making process.\nSaliency methods [4, 5, 6, 7, 8, 9] have become a popular tool to explain predictions of a trained\nmodel by highlighting parts of the input that presumably have a high relevance for its predictions.\nHowever, to the best of our knowledge the existing saliency methods are mostly focused on image\nclassi\ufb01cation networks and thus are not able to spatially differentiate between prediction explanations.\nIn this work, we propose a way to extend existing saliency methods designed for image classi\ufb01cation\ntowards (pixel-level) dense prediction tasks, which allows to generate spatially coherent explanations\nby exploiting spatial information in dense predictions. We call our approach grid saliency, which\nis a perturbation-based saliency method, formulated as an optimization problem of identifying the\nminimum unperturbed area of the image needed to retain the network predictions inside a target object\n\n33rd Conference on Neural Information Processing Systems (NeurIPS 2019), Vancouver, Canada.\n\n\f(b) Context explanations (red) for different segments (outlined in blue)\n\n(a) Semantic segment.\nFigure 1: Context explanations by grid saliency for semantic segmentation [1, 2] on Cityscapes [3].\nGrid saliency not only can contextually explain correct predictions: in the \ufb01rst row the network\nlooks at the motorbike to correctly predict the class rider (light blue); but can also explain erroneous\npredictions: in the second row the upper body of the rider is incorrectly predicted as person, but for\nthis prediction the bicycle is not salient in contrast to the correctly predicted legs of the rider.\n\nregion. As our grid saliency allows to differentiate between objects and their associated context areas\nin the saliency map, we speci\ufb01cally explore its potential to produce context explanations for semantic\nsegmentation networks. The contextual information is known to be one of the essential recognition\ncues [10, 11, 12], thus we aim to investigate which local and global context is the most relevant for\nthe network class predictions inside a target object area (see Fig. 1 and Fig. 2 for examples).\nIn real-world scenarios, context biases are inherently present in the data. To evaluate whether the\nproposed grid saliency is sensitive to context biases and has the ability to detect them, we introduce a\nsynthetic toy dataset for semantic segmentation, generated by combining MNIST digits [13] with\ndifferent fore- and background textures, for which we arti\ufb01cially induce a context bias between the\ndigit and the background texture. Besides detecting the mere presence of the context bias, we also\nanalyze the ability of our method to localize it in the image (see Sec. 4). We employ this dataset to\ncompare our approach with different baselines, i.e., introduced extensions of gradient-based saliency\nmethods of [4, 14, 15] to produce context explanations. We show that the proposed dataset can serve\nas a valid benchmark for assessing the quality of saliency methods to detect and localize context\nbiases. We \ufb01nd that gradient-based techniques, in contrast to our grid saliency, are ill-suited for the\ncontext bias detection. By design they tend to produce noisy saliency maps which are not faithful to\nthe context bias present in the data, whereas our method has higher sensitivity for context bias and\nthus can also precisely localize it in the image. We further evaluate grid saliency performance to\nproduce context explanations for semantic segmentation on the real-world Cityscapes dataset [3] and\nexperimentally show that the produced context explanations faithfully re\ufb02ect spatial and semantic\ncorrelations present in the data, which were thus picked up by the segmentation network [1, 2].\nTo the best of our knowledge, we are the \ufb01rst to extend saliency towards dense-prediction tasks and\nuse it to produce context explanations for semantic segmentation. In summary, our contributions are\nthe following: (1) We propose an extension of saliency methods designed for classi\ufb01cation towards\ndense prediction tasks. (2) We exploit the proposed grid saliency to produce context explanations\nfor semantic segmentation and show its ability to detect and localize context biases. (3) We create a\nsynthetic dataset to benchmark the quality of produced explanations as well as their effectiveness for\ncontext bias detection/localization. (4) We investigate the faithfulness of context explanations for\nsemantic segmentation produced by the grid saliency on real-world data.\n\n2 Related Work\nExplanations. Many methods attempt to interpret the network decision making process by pro-\nducing explanations via bounding boxes [16, 17] or attributes [18], providing textual justi\ufb01cations\n[19, 20, 21] or generating low-level visual explanations [4, 22, 5, 6]. Our work builds on top of\nthe latter approaches, also known as saliency methods, which try to identify the image pixels that\ncontribute the most to the network prediction. These methods mostly focus on the task of image\nclassi\ufb01cation and can be divided into two categories: gradient-based and perturbation-based methods.\nGradient-based methods [24, 4, 25] compute a saliency map that visualizes the sensitivity of each\nimage pixel to a speci\ufb01c class prediction, which is obtained by backpropagating the gradient for the\nprediction with respect to the image and estimating how moving along the gradient in\ufb02uences the\nclass output. To circumvent noise and visual diffusion in saliency maps, [14] proposes to sum up the\n\n2\n\n\f(b) Semantic segm.\n\n(c) Image w/o context\n\n(a) Context explanation\nFigure 2: Effect of context on semantic segmentation and the context explanation provided by grid\nsaliency for an erroneous prediction, the image is taken from MS COCO [23]. The grid saliency (a)\nshows the responsible context for misclassifying the cow (green) as horse (purple) in the semantic\nsegmentation (b). It shows the training bias that horses are more likely on road than cows. Removing\nthe context (c) yields a correctly classi\ufb01ed cow (d).\n\n(d) Segm. w/o context\n\ngradients over the intensity-scaled input versions, while [15] averages over many noisy samples of\nthe input. Other methods [5, 26, 27] explore integrating network activations into their saliency maps.\nGradient-based methods mostly rely on heuristics for backpropagation and as has been shown by [7]\nmay provide explanations which are not faithful to the model or data.\nPerturbation-based methods [6, 28, 9, 8] evaluate the class prediction change with respect to a\nperturbed image, e.g. in which speci\ufb01c regions of the image are either replaced with the mean image\nvalues or removed by applying blur or Gaussian noise. The approach in [6] formulates this output\nchange as an optimization problem of the original and perturbed image, while [28] estimates the\nperturbation mask by training an auxiliary network. To account for image discontinuities, [9, 8]\ncompute the saliency of the masked region by marginalizing it out either over neighboring image\nregions or by conditioning the trained generative model on the non-masked region and then estimating\nthe classi\ufb01cation change. Perturbation approaches might be vulnerable to network artifacts resulting\nin arbitrary saliency regions [8]. To overcome this, [6, 28] resort to perturbe large image regions.\nThe above methods are limited to explanations of image classi\ufb01cation. In this work, we propose a\nway to extend them to produce grid saliency maps for dense prediction networks as well (see Sec. 3).\nWe showcase the usability of grid saliency for context explanations of semantic segmentation (see\nSec. 4 and 5). The closest related work to produce context explanations is [29], which proposes a\nnetwork that jointly learns to predict and contextually explain its decisions. In contrast to [29], we\nfocus on the post hoc model explanations and are not limited to the image classi\ufb01cation task.\n\nSemantic segmentation. CNNs have become a default technique for semantic segmentation [30,\n31, 32]. [33] \ufb01rst showcased the use of CNNs for segmentation. Since then, multiple techniques\nhave been proposed, from utilizing dilated convolutions [34, 35] and post-processing smoothing\noperations [36, 37, 38, 39] to employing spatial pyramid pooling [40, 30, 41, 2] and encoder/decoder\narchitectures [42, 43, 44, 45]. As the accuracy of these methods comes at a high computational cost,\nthere has been an increasing interest in developing real-time semantic segmentation networks with\nlow memory needs [46, 1, 47, 48, 49]. In this work, we aim to produce context explanations for\nsemantic segmentation. To investigate the effectiveness of our approach and show its generalization\nacross architectures, we employ DeepLabv3+ [2] and U-Net [44] with different backbones [1, 50, 51].\n\n3 Method\nIn Sec. 3.1 we introduce the grid saliency method, which allows to produce spatially coherent\nexplanations for dense predictions, and present a way to use it for context explanations of semantic\nsegmentation. Next, in Sec. 3.2 we extend the popular gradient-based saliency methods [4, 14, 15] to\nproduce spatial explanations as well, which we later compare with grid saliency in Sec. 4.\n\n3.1 Grid Saliency via Perturbation\nLet f : I \u2192 O denote the prediction function, e.g. a deep neural network, which maps a grid\ninput space I = RHI\u00d7WI\u00d7CI to a grid output space O = RHO\u00d7WO\u00d7CO, where W and H are\nthe respective width and height of the input and output, and CI and CO are the number of input\nchannels (e.g. 3 or 1 for images) and output prediction channels (e.g. number of classes for semantic\nsegmentation). To keep the discussion concrete, we consider only images as input, x \u2208 I, and\nper-pixel dense predictions of the network f (x) \u2208 O as output. The goal is to \ufb01nd the smallest\nsaliency map M \u2208 [0, 1]HI\u00d7WI that must retain in the image x in order to preserve the network\n\n3\n\n\fprediction in the request mask area R \u2208 {0, 1}HO\u00d7WO for class (channel) c \u2208 {1, ..., CO}. Further\non, for simplicity we assume that the input and output spatial dimensions are the same.\nOur method builds on top of perturbation saliency methods [6, 28, 9, 8] designed for image classi\ufb01ca-\ntion. They aim to \ufb01nd salient image regions most responsible for a classi\ufb01er decision by replacing\nparts of the image with uninformative pixel values, i.e. perturbing the image, and evaluating the\ncorresponding class prediction change. We follow the same image perturbation strategy as [6]. Let p\ndenote a perturbation function that removes information from an image x outside of the saliency M.\nFor example, such perturbation function can be the interpolation between x and a \u2208 I, where a can be\na constant color image, gaussian blur, or random noise. In this case, p(x, M ) = x\u25e6 M + a\u25e6 (1\u2212 M ).\nNote, that in practice M operates on a lower resolution to avoid adversarial artifacts [6, 28] and is\nlater upsampled to the original image resolution. In addition, the pixel values of the perturbed image\np(x, M ) are clipped to preserve the range of the original image space.\nWith this notation in hand, we can formulate the problem of \ufb01nding the saliency map M for the\nprediction of class c as the following optimization problem:\n\n\u03bb \u00b7 (cid:107)M(cid:107)1 + (cid:107) max(fc(x) \u2212 fc(p(x, M )), 0)(cid:107)1,\n\n(1)\nwhere (cid:107) \u00b7 (cid:107)1 denotes the l1 norm and fc(x) is the network prediction for class c . The \ufb01rst term can\nbe considered as a mask loss that minimizes the salient image area and perturbs the original image\nas much as possible. The second term serves as a preservation loss which ensures that the network\nprediction fc(p(x, M )) for class c on the perturbed image p(x, M ) reaches at least the con\ufb01dence\nof the network prediction fc(x) on the original unperturbed image. Thus, the second loss term can\nbe considered as a penalty for not meeting the constraint fc(p(x, M )) (cid:62) fc(x), hence the use of\nmax(\u00b7, 0) in Eq. 1. The parameter \u03bb controls the sparsity of M.\nWe then can spatially disentangle explanations given in the saliency map M for the network predic-\ntions in the requested area of interest R from the explanations for the other predictions, by restricting\nthe preservation loss to the request mask R in Eq. (1):\n\nM\n\n(x, c) = argmin\n\nM\n\n\u2217\n\n\u2217\ngrid(x, R, c) = argmin\n\nM\n\n\u03bb \u00b7 (cid:107)M(cid:107)1 +\n\nM\n\n(cid:107)R \u25e6 max(fc(x) \u2212 fc(p(x, M )), 0)(cid:107)1\n\n.\n\n(2)\n\n(cid:107)R(cid:107)1\n\ngrid in Eq. (2) as a grid saliency map.\n\nFurther on, we will refer to M\u2217\nContext explanations for semantic segmentation. We now adapt the grid saliency formulation\nfrom Eq. (2) to speci\ufb01cally provide context explanations for the requested area of interest R. Context\nexplanations are of particular interest for semantic segmentation, as context often serves as one of\nthe main cues for semantic segmentation networks (see Fig. 2). Thus, here we focus on context\nexplanations for semantic labelling predictions and assume that R is the area covering the object\nof interest in the image x. To optimize for salient parts of the object context, we integrate the\nobject request mask R in the perturbation function. For the request mask R, the perturbed image\np(x, R) \u2208 I contains only the object information inside R and all the context information outside R\nis removed (with a constant color image a). For optimization, we will now use this new perturbed\nimage p(x, R) instead of the maximally perturbed image p(x, M = 0) = a and denote the context\nperturbation function as pcontext(x, R, M ) = x \u25e6 M + p(x, R) \u25e6 (1 \u2212 M ).\nThe context saliency map for class c and request object R can be obtained via optimization of\n\nM\n\n\u2217\ncontext(x, R, c) = argmin\n\nM\n\n\u03bb \u00b7 (cid:107)M(cid:107)1 +\n\n(cid:107)R \u25e6 max(fc(x) \u2212 fc(pcontext(x, R, M )), 0)(cid:107)1\n\n,\n\n(3)\n\n(cid:107)R(cid:107)1\n\nwhere the saliency map is optimized to select the minimal context necessary to at least yield the\noriginal prediction for class c inside the request mask R. Note that, M\u2217\ncontext can be an empty mask if\nno context information is needed to recover the original prediction inside R.\n\n3.2 Gradient-Based Variants\n\nAnother way to produce spatially coherent saliency maps is to make use of the popular gradient-based\nsaliency methods. Thus, we additionally consider the Vanilla Gradient (VG) [24], Integrated Gradient\n(IG) [14], and SmoothGrad (SG) [15] saliency methods.\nLet G(x, c) = \u2202gc(x)/\u2202x \u2208 RHI\u00d7WI\u00d7CI denote the gradient of the classi\ufb01cation network prediction\ngc(x) \u2208 R for class c with respect to the input image x \u2208 I. For the classi\ufb01cation task, the saliency\n\n4\n\n\fM VG(x, c) =\n\n(cid:12)(cid:12)G(x, c)(cid:12)(cid:12), M SG(x, c) =\n\n(cid:88)\n\nc\u2208CI\n\nmaps of VG, IG and SG are computed as:\n\n(cid:12)(cid:12)(cid:12)(cid:12) 1\n\nn\n\n(cid:88)\n\nc\u2208CI\n\nG(cid:0)x + N (0, \u03c32), c(cid:1)(cid:12)(cid:12)(cid:12)(cid:12),\nn(cid:88)\n(cid:12)(cid:12)(cid:12)(cid:12) 1\n(cid:19)(cid:12)(cid:12)(cid:12)(cid:12),\n(cid:88)\nn(cid:88)\n\n(cid:18) k\n\nk=1\n\nx, c\n\nn\n\nG\n\nc\u2208CI\n\nn\n\nk=1\n\n(4)\n\nM IG(x, c) =\n\nwhere n is the number of approximation steps for IG or the number of samples for SG, and N (0, \u03c32)\nrepresents Gaussian noise with standard deviation \u03c3.\nFollowing Sec. 3.1, we next extend the above approaches to produce the saliency M for dense\npredictions fc(x) \u2208 O and to spatially disentangle explanations given in the saliency M for the\nnetwork predictions in the request area R from other predictions. For a given input x and a binary\nrequest mask R, we denote the normalized network prediction score for class c in the request area R as\nS(x, R, c) = (cid:107)R \u25e6 fc(x)(cid:107)1/(cid:107)R(cid:107)1, S(x, R, c) \u2208 R. Similarly to G(x, c), we de\ufb01ne Ggrid(x, R, c) ..=\n\u2202S(x, R, c)/\u2202x \u2208 RHI\u00d7WI\u00d7CI which directly yields M VG/SG/IG\n(x, R, c) by replacing G(x, c) in\nEq. 4 with Ggrid(x, R, c). For the gradient-based context saliency, as in Sec. 3.1 only salient pixels\noutside of the object area are considered, i.e.\n\ngrid\n\nM VG/IG/SG\n\ncontext\n\n(x, R, c) ..= (1 \u2212 R) \u25e6 M VG/IG/SG\n\ngrid\n\n(x, R, c).\n\n(5)\n\nGradient-based saliency maps are prone to be noisy. Thus, to circumvent this and also make them\nmore comparable to the lower resolution perturbation-based grid saliency, in our experiments we\napply the spatial mean \ufb01lter on top of the saliency map with a (WI /WS) \u00d7 (HI /HS) kernel and\nstride, where WS \u00d7 HS is the resolution of the perturbation-based saliency map.\n\n4 Context Bias Detection on Synthetic Data\nAlthough context biases are inherently present in real-world data, in practice it is hard to measure\nand alter these correlations and thus to employ this data for benchmarking context bias detection. In\norder to evaluate whether the grid saliency is sensitive to context biases and has the ability to detect\nthem, in Sec. 4.1 we introduce a synthetic dataset for semantic segmentation with arti\ufb01cially induced\ncontext biases.1 We use this dataset to compare the grid saliency methods proposed in Sec. 3 and\nshow that this dataset can serve as a valid benchmark for assessing the quality of saliency methods\nfor context bias detection and localization, see Sec. 4.2.\n\n4.1 Benchmark for Context Bias Detection\nDataset. The proposed synthetic toy dataset consists of gray scale images of size 64 \u00d7 64 pixels,\ngenerated by combining upscaled digits from MNIST [13] with foreground and background textures\nfrom [52, 53], as can be seen in Fig. 3 (a). In order to introduce different context information for the\ndigit, two background textures are used for the upper and lower half of the image. For each synthetic\nimage a corresponding segmentation ground truth is generated, where the MNIST digit de\ufb01nes the\nmask and the semantic class (including a background class).\nTo evaluate the ability of saliency methods to explain context biases, we propose to generate biased\nand unbiased versions of the dataset. For the unbiased version (DSno-bias), all fore- and background\ntextures appear with equal probability for all digits. For the biased version, a single digit class is\ncoupled with a speci\ufb01c background texture, located randomly either in the upper or lower half of\nthe background. We consider two variants of it, with a weakly DSw-bias and strongly induced bias\nDSs-bias. For the dataset with a strongly induced bias DSs-bias, a speci\ufb01c texture appears if and only\nif a certain digit class is present. For a weakly induced bias DSw-bias, a speci\ufb01c texture always appears\nalong with the biased digit but also uniformly appears with other digits. From the pool of 25 textures,\n5 textures are chosen randomly to induce context bias for one of 10 digits. For all 50 texture/digit\ncombinations, a weakly and strongly biased dataset variant with train/test splits is generated.\n\n1The code for generating the proposed synthetic dataset with induced context biases can be found here:\n\nhttps://github.com/boschresearch/GridSaliency-ToyDatasetGen.\n\n5\n\n\fFigure 3: a) Synthetic dataset images, see Sec. 4.1 for details. b) Context saliency maps of different\nmethods. For the image on the left, the different context saliency maps Mcontext are shown for the\nnetworks [1] trained on the unbiased (top row) and biased (bottom row) dataset versions. For the\nbiased version, the digit 2 is biased with the background texture in the top half of the image. Our\nperturbation-based grid saliency is able to precisely detect this bias, in contrast to other methods.\nEvaluation metrics. To evaluate the extent to which a network is able to pick up a context bias\npresent in the training data, we propose to measure on the unbiased DSno\u2212bias test set the performance\nof the network trained on the biased DSw/s-bias training set, using the standard Intersection over\nUnion (IoU, see [33]) metric for semantic segmentation. If the network has picked up the bias, we\nexpect to see a signi\ufb01cant drop in IoU for the biased digit segmentation.\nTo benchmark how well different saliency methods (perturbation- or gradient-based) can detect the\ncontext bias of semantic segmentation networks, we propose to evaluate to which extent a context\nsaliency map Mcontext(x, R, c) (cf. Eq. 3 and 5 in Sec. 3) for the request object mask R and its\ncorresponding class c is concentrated on the ground truth context area C = 1 \u2212 RGT , where RGT is\nthe ground truth mask of the object R, by using a context bias detection metric (CBD):\n\nCBD(x, R, c) = (cid:107)C \u25e6 Mcontext(x, R, c)(cid:107)1/(cid:107)C(cid:107)1.\n\n(6)\nTo benchmark the ability of different saliency methods to localize a context bias, we propose to\nmeasure how much of the context saliency Mcontext(x, R, c) falls into the ground truth biased context\narea Cbias, the upper or lower half of the image x by the design of DSw/s-bias, where Cbias is a binary\nmask of the biased context area. We refer to this metric as a context bias localization (CBL):\n\nCBL(x, R, c) = (cid:107)Cbias \u25e6 Mcontext(x, R, c)(cid:107)1/(cid:107)C \u25e6 Mcontext(x, R, c)(cid:107)1.\n\n(7)\nIn our experiments we report mIoU and the mean CBD and CBL measures (mCBD, mCBL) per\nbiased digit class c, averaging the results across all images in the DSw/s-bias test set, 5 randomly\nselected bias textures and 5 different initial random seeds for the training set generation.\n\n4.2 Experimental Results\nImplementation details. We use the U-Net [44] architecture with a VGG16 [51] backbone. As\nrequest mask, the segmentation prediction of the target digit is used. The saliency maps with a size of\n4\u00d7 4 are optimized using SGD with momentum of 0.5 and a learning rate of 0.2 for 100 steps starting\nwith a 0.5 initialized mask. A weighting factor of \u03bb = 0.05 is used (see Eq. 3). A constant color\nimage is used for perturbation. Further implementation details are provided in the supp. material.\nBias in the trained network. We \ufb01rst investigate if the networks trained on DSw-bias and DSs-bias\nhave picked up the induced weak and strong context biases in the training data. For this purpose, we\nevaluate their performance in terms of mIoU on the unbiased dataset DSno-bias and report the results\nin Fig. 4 (a), which visualizes the digit-wise mIoU with respect to the biased digit. The \ufb01rst row of\nthe heat map (labeled as N) in Fig. 4 (a) shows the performance of the networks trained on DSno-bias.\nWe observe a clear drop in performance for biased digits (diagonal elements) in comparison to the\n\ufb01rst row. As expected, the performance drop is higher for the stronger bias. Moreover, the mIoU\nof the unbiased digits (non-diagonal elements) is also affected by the introduced bias. For example,\ninducing a bias for the digit nine leads to a decreased performance for the digit four (see second row\nin Fig. 4 (a)). We observe that this effect mostly occurs for the similar looking digits and, most likely,\nis caused by the fact that on the unbiased dataset the bias textures also occur with the unbiased digits,\nresulting in the confusion of similar looking digits for the network. From the observed mIoU drop for\nbiased digits we can conclude that the networks have picked up the introduced bias. However, in real\nworld it is often impossible to collect fully unbiased data. Thus, we next evaluate the ability of our\ngrid saliency to detect context bias only using the biased data.\n\n6\n\nToy data examplesa)InputVGOursb)UnbiasedBiasedIGSG\f(a) Segmentation\n\n(b) Bias detection\n\n(c) Bias localization\n\nFigure 4: Context bias detection/localization of different saliency methods. (a) shows the segmenta-\ntion mIoU and (b) & (c) report the context bias detection and localization results, for the strongly and\nweakly biased datasets, respectively. In (a)-(b), the y axis indicates the network bias towards the digit\n(0-9), with N denoting the unbiased setting. The x axis indicates the corresponding digit result. In\ncontrast to VG [24], IG [14], and SG [15], our grid saliency is able to accurately detect (see diagonal\nelements in (b)) and localize (c) the induced context bias, see Sec. 4.2 and suppl. material for details.\nContext bias detection with grid saliency.\nIn the last column of Fig. 4 (b), we report the context\nbias detection results for our perturbation-based grid saliency method, described in Sec. 3.1, using\nthe CBD metric (see Eq. 6). The mCBD values are visualized with respect to networks trained on\ndata biased to different digits DSs/w-bias (y-axis) and for the different digit classes (x-axis) in the\nbiased test set. The only exception is the \ufb01rst row (labeled as N), where for comparison we show the\nresults with no bias, for the network trained on DSno-bias. We observe that our grid saliency shows\nsubstantial evidence of context bias for digits with induced bias (diagonal elements), both strong and\nweak. Even for the weak bias in Fig. 4 (b) the grid saliency still clearly differentiates between biased\nand unbiased digits (diagonal vs. non-diagonal elements). Note that the bias detection using the grid\nsaliency does not require an unbiased test set. In the suppl. material, we also study the in\ufb02uence of\nthe bias texture as well as the choice of hyperparameters.\nComparison across different saliency methods.\nIn Fig. 4 (b) and (c) we compare our perturbation-\nbased grid saliency with the context saliency extensions of gradient-based methods, i.e. VG, SG, and\nIG (see Eq. 4), using mCBD and mCBL metrics proposed in Sec. 4.1. From Fig. 4 (b) we see that VG\nand SG are not able to reliably detect the context bias, while IG achieves a comparable performance\nto our grid saliency. However, in contrast to the perturbation saliency, IG has also high mCBD values\nfor unbiased digits, complicating its use in practice, as one would need to tune a reliable detection\nthreshold, which is particularly challenging for weakly biased data.\nContext bias localization.\nIn addition to the detection performance, we evaluate how well the\nsaliency methods localize the context bias using the mCBL measure. Fig. 4 (c) shows that VG, IG,\nand SG have low localization performance, comparable to random guessing (\u223c 0.5 mCBL), while\nour grid saliency is the only method which is able to accurately localize the context bias (mCBL\nabove 0.9) on both strongly and weakly biased data. We evaluate the ability of our perturbation\ngrid saliency to detect and localize context bias across different semantic segmentation networks.\nFor U-Net [44] with the VGG16 [51], ResNet18 [50] and MobileNetv2 [1] backbones, we observe\nsimilar bias detection and localization performance (see also supp. material).\n\n5 Cityscapes Experiments\nMotivated by the success of our perturbation-based grid saliency at detecting context biases on the\nsynthetic dataset described in Sec. 4, we next apply it to Cityscapes [3] in order to produce and\nanalyze context explanations for semantic segmentation in real-world scenes.\nExperimental setup. We use 500 \ufb01nely annotated images of the Cityscapes validation set, con-\nsidering only a subset of classes for the analysis. For our experiments we use the Deeplabv3+ [2]\nnetwork with a Mobilenetv2 backbone [1]. Our optimization setup largely carries over from Sec. 4.2,\nwith the exception that we optimize a coarse 16 by 32 pixel mask using SGD with a learning rate of 1\nfor 80 steps and use \u03bb = 0.01. Additional implementation details are provided in the supp. material.\n\n7\n\nN9876543210StrongbiasmIoU0123456789DigitN9876543210WeakbiasVGSGIGOurs0123456789Digit0123456789Digit0123456789Digit0123456789Digit0.00.51.0mCBL(strong)VGSGIGOurs0.00.51.0mCBL(weak)\fSemantic segment. [40]\n\nPerson (corr)\n\nRider (err)\n\nRider (err)\n\nRider (corr)\n\nPerson (corr)\n\nPole (corr)\n\nTraf\ufb01c sign (corr)\n\nFigure 5: Context explanations by grid saliency for erroneous (err) and correct (corr) semantic\nsegmentation predictions on Cityscapes [3]. Salient red regions visualize the most important context\nMcontext for the requested segment prediction R outlined in blue. See Sec. 5 for discussion.\n\nFigure 6: Class statistics of context explanations on Cityscapes, see Sec. 5 for discussion.\n\nAnalysis of context explanations.\nIn order to gain a global understanding of the model context\nbias, we aggregate statistics from the produced grid saliency maps for several requested classes across\nall validation images. These statistics are summarized in Fig. 6. For each image and class, the context\nsaliency is computed for all suf\ufb01ciently large instances (at least 10k pixels), as can be seen in the\nsecond row of Fig. 5. Next, for each image and semantic class we compute the weighted label\ndistribution of context salient pixels. The weight of each pixel is given by the computed saliency\nintensity at said pixel and its context class label is taken from the ground truth segmentation. For each\nrequested object class, we group by label across all images and sum the saliency values within each\ngroup. This yields an accumulated saliency value for each label. We normalize the results by the sum\nof all saliency values across all classes and images to yield a probability distribution across labels.\nTo show that context explanations produced by our grid saliency are meaningful, we additionally\ncompare its accumulated context class statistics with their baseline class distributions. These baselines\nare computed in the same manner as above, but instead of relying on the optimized saliency map we\nuse a \ufb01xed dilation of the contour of all predicted object segments considered in the image, thereby\ncapturing the immediate context around each object in a uniform set of directions.\nFig. 6 compares these accumulated statistics of context explanations for the requested object classes,\ngiven on the y axis. We observe that the context explanations are focused across reasonable and\nsomewhat expected class subsets, which vary per class. For instance, in comparison to the baseline,\ncontext saliency for the rider class shifts attention from its spatially co-occurring classes such as\nroad, vegetation and building, to mainly the bicycle class. Car context saliency attention is decreased\non building and vegetation and mainly focused on road, sidewalk and other cars. Bicycle context\nsaliency mostly attends to sidewalk rather than the road class. Overall, our grid saliency is able to\nprovide sensible and coherent explanations for network decision making, which re\ufb02ect semantic\ndependencies present in street scenes. These quantitative \ufb01ndings are validated by the qualitative\nresults in Fig. 5. The second row of Fig. 5 gives several representative examples of context saliencies.\nPlease refer to the supp. material for additional examples and a high-resolution version of Fig. 6 as\nwell as a qualitative model comparison on MS COCO.\nContext explanations of erroneous predictions. One of the motivations for this work is also to\nexplain unexpected model behavior. Speci\ufb01cally, in case of erroneous predictions (i.e., Fig. 1 and 2),\nwe wish to understand to which extent context bias contributes to the failure. Table 1 shows class\nstatistics of context explanations for correct and erroneous predictions, where the intersection of the\nground truth mask and the correct or misclassi\ufb01ed prediction region are used as the request mask\nR, and similarly to Fig. 6 semantic class statistics are computed inside the salient context for all\nsuf\ufb01ciently large request masks (at least 625 pixels). We observe that the saliency severely changes\n\n8\n\nroadbuildingvegetationsidewalkcarbicyclemotorcyclepersonriderpoletraffic_signContext Classcarbicyclemotorcyclepersonriderpoletraffic_signRequest Classa) Grid Saliencyroadbuildingvegetationsidewalkcarbicyclemotorcyclepersonriderpoletraffic_signContext Classb) Baseline0.000.080.160.240.320.40roadbuildingvegetationsidewalkcarbicyclemotorcyclepersonriderpoletraffic_signContext ClassDifference of (a) and (b)0.160.080.000.080.16\fTable 1: Class statistics of context explanations of correct and erroneous predictions on Cityscapes.\n\nContext class\n\nGT\n\nrider\n\nperson\n\nbicycle\n\ncar\n\nmotorcycle\n\npole\n\nPrediction\n\nrider\nperson\nperson\nrider\ncar\n\nbicycle\n\ncar\n\nroad bicycle motorcycle car vegetation building sidewalk rider person pole\n0.07 0.02 0.02\n0.17 0.22\n0.08 0.11 0.04\n0.13\n0.10\n0.12\n0.03\n0.01 0.05 0.05\n0.08 0.15 0.03\n0.09 0.18\n0.01 0.19 0.01\n0.01\n0.19\n0.05 0.04 0.04\n0.18\n0.06\nmotorcycle 0.19 0.27\n0.03 0.01 0.04\n0.39 0.02\n0.01 0.04 0.04\n0.01 0.00 0.04\n0.06\n0.01 0.22 0.02\n0.01\nmotorcycle 0.18 0.04\n0.05 0.03 0.06\n0.10\n0.01 0.05 0.03\n0.05 0.04 0.06\n0.01\n0.00 0.03 0.04\n0.02\n0.00 0.01 0.05\n0.01\n0.00\n0.00 0.00 0.07\n\n0.07\n0.05\n0.09\n0.04\n0.32\n0.05\n0.01\n0.06\n0.17\n0.28\n0.04\n0.10\n0.10\n0.05\n0.04\n0.03\n\n0.11\n0.24\n0.22\n0.13\n0.11\n0.15\n0.11\n0.10\n0.07\n0.14\n0.14\n0.14\n0.10\n0.23\n0.57\n0.08\n\n0.09\n0.10\n0.15\n0.10\n0.05\n0.09\n0.02\n0.11\n0.14\n0.03\n0.07\n0.10\n0.10\n0.28\n0.07\n0.63\n\n0.10\n0.05\n0.13\n0.07\n0.03\n0.19\n0.13\n0.08\n0.02\n0.07\n0.18\n0.11\n0.11\n0.10\n0.02\n0.03\n\nbicycle\n\ncar\npole\n\n0.13\n0.16\n0.04\nbuilding\n0.04\nvegetation 0.03\n\n0.03\n0.01\n0.00\n0.02\n0.00\n0.01\n0.13\n0.00\n0.22\n0.00\n0.10\n0.07\n0.18\n0.00\n0.00\n0.00\n\nmotorcycle 0.07\n0.14\n\nperson\n\nfor the error cases in comparison to the correct predictions. E.g. for the correctly classi\ufb01ed rider\nclass, context saliency mostly focuses on bicycle (22%), but is much less present (10%) when rider is\nmistaken as person and in this case the context saliency shifts from bicycle to building (24%). The\nopposite effect is observed when person is misclassi\ufb01ed as rider. Sidewalk saliency is decreased when\nbicycle is misclassi\ufb01ed as motorcycle (19% vs. 13%) and when pole is misclassi\ufb01ed as building (10%\nvs. 2%). In general, grid saliency is able to provide reasonable explanations of erroneous predictions.\nFig. 1 (second row) and Fig. 5 (\ufb01rst row) showcase how applying grid saliency on the same input\nimage but different prediction outputs is useful for isolating failures caused by context bias. Both\nof them show examples of context explanations for erroneous segmentations of a single object. In\nFig. 5, the arms, head and shoulders of a pedestrian are classi\ufb01ed as rider, while the rest of the body\nis correctly identi\ufb01ed as person. The context saliency for the former body parts activates highly on\nthe arms and stroller handles, whereas activations for person does not highlight this support. Thus, a\nreasonable conclusion is that the misclassi\ufb01cation may be attributed to the arm pose and potentially\nalso to the context bias given by the similar appearance of the stroller and bicycle handles.\n6 Conclusions\nWe proposed spatial grid saliency, a general framework to produce spatially coherent explanations\nfor (pixel-level) dense prediction networks, which to the best of our knowledge is the \ufb01rst method to\nextend saliency techniques beyond classi\ufb01cation models. We investigated the ability of grid saliency\nto provide context explanations for semantic segmentation, showing its effectiveness to detect and\nlocalize context bias on the synthetic toy dataset speci\ufb01cally designed with an arti\ufb01cially induced\nbias to benchmark this task. Our results on the real-world data indicated that grid saliency can be\nsuccessfully employed to produce easily interpretable and faithful context explanations for semantic\nsegmentation, helping to discover spatial and semantic correlations in the data picked up by the\nnetwork. We hope the proposed grid saliency and the insights of this work can contribute to a better\nunderstanding of semantic segmentation networks or other models for dense prediction, elucidating\nsome aspects of the problem that have not been well explored so far.\nBesides enabling visual explanations for dense-prediction tasks, we see potential utility of grid\nsaliency for the following applications: 1) Architecture comparison: Context explanations produced\nby grid saliency can be used to compare architectures with respect to their capacity to either learn\nor to be invariant towards context. 2) Network generalization via active learning: Existing context\nbiases might impair network generalization. E.g., cows might mostly appear on grass during training.\nA network that was trained and evaluated on this data and picked up that bias will perform poorly in\nreal-world cases, where the cow, for example, appears on road and gets misclassi\ufb01ed as the horse\nclass, see Fig. 2. Actions can be taken, such as targeted extra data collection, to improve network\ngeneralization. 3) Adversarial detection: Grid saliency can be used to detect and localize adversarial\npatches outside object boundaries (e.g. [54]). Cases for which the salient regions lie largely outside\nan object would strongly indicate the presence of an adversary or misguided prediction. We consider\nthe above-mentioned utilities of grid saliency interesting and promising future research directions.\n\n9\n\n\fReferences\n[1] Mark B. Sandler, Andrew G. Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen.\nIn Proceedings of the IEEE Conference on\n\nMobilenetv2: Inverted residuals and linear bottlenecks.\nComputer Vision and Pattern Recognition (CVPR), pages 4510\u20134520, 2018.\n\n[2] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder\nwith atrous separable convolution for semantic image segmentation. In Proceedings of the European\nConference on Computer Vision (ECCV), 2018.\n\n[3] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson,\nUwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding.\nIn Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016.\n\n[4] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising\nimage classi\ufb01cation models and saliency maps. In International Conference on Learning Representations\n(ICLR), 2013.\n\n[5] Ramprasaath R. Selvaraju, Abhishek Das, Ramakrishna Vedantam, Michael Cogswell, Devi Parikh, and\n\nDhruv Batra. Grad-cam: Why did you say that? arXiv:1611.07450, 2016.\n\n[6] Ruth Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In\nProceedings of the IEEE International Conference on Computer Vision (ICCV), pages 3449\u20133457, 2017.\n[7] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian J. Goodfellow, Moritz Hardt, and Been Kim. Sanity\n\nchecks for saliency maps. In Advances in Neural Information Processing Systems (NeurIPS), 2018.\n\n[8] Chun-Hao Chang, Elliot Creager, Anna Goldenberg, and David Duvenaud. Explaining image classi\ufb01ers by\n\ncounterfactual generation. In International Conference on Learning Representations (ICLR), 2019.\n\n[9] Luisa M. Zintgraf, Taco Cohen, Tameem Adel, and Max Welling. Visualizing deep neural network\ndecisions: Prediction difference analysis. In International Conference on Learning Representations (ICLR),\n2017.\n\n[10] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel\nUrtasun, and Alan Loddon Yuille. The role of context for object detection and semantic segmentation in\nthe wild. 2014.\n\n[11] Jasper RR Uijlings, Arnold WM Smeulders, and Remko JH Scha. The visual extent of an object.\n\nInternational journal of computer vision, 96(1):46\u201363, 2012.\n\n[12] Aymen Azaza, Joost van de Weijer, Ali Douik, and Marc Masana. Context proposals for saliency detection.\n\nComputer Vision and Image Understanding, 174:1\u201311, 2018.\n\n[13] Yann LeCun, Corinna Cortes, and Christopher J.C. Burges. The MNIST database of handwritten digits.\n\nTechnical report, 1998.\n\n[14] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In International\n\nConference on Machine Learning (ICML), 2017.\n\n[15] Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda B. Vi\u00e9gas, and Martin Wattenberg. Smoothgrad:\n\nremoving noise by adding noise. arXiv:1706.03825, 2017.\n\n[16] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In\n\nProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015.\n\n[17] M. Oramas Jos\u00e9Oramas and Tinne Tuytelaars. Modeling visual compatibility through hierarchical mid-level\n\nelements. arXiv:1604.00036, 2016.\n\n[18] Sadaf Gulshad, Jan Hendrik Metzen, Arnold W. M. Smeulders, and Zeynep Akata. Interpreting adversarial\n\nexamples with attributes. arXiv:1904.08279, 2019.\n\n[19] Dong Huk Park, Lisa Anne Hendricks, Zeynep Akata, Anna Rohrbach, Bernt Schiele, Trevor Darrell,\nand Marcus Rohrbach. Multimodal explanations: Justifying decisions and pointing to the evidence. In\nProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018.\n\n[20] Lisa Anne Hendricks, Zeynep Akata, Marcus Rohrbach, Jeff Donahue, Bernt Schiele, and Trevor Darrell.\nGenerating visual explanations. In Proceedings of the European Conference on Computer Vision (ECCV),\n2016.\n\n[21] Yinpeng Dong, Hang Su, Jun Zhu, and Bo Zhang. Improving interpretability of deep neural networks with\nsemantic information. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition\n(CVPR), 2017.\n\n[22] Bolei Zhou, Aditya Khosla, \u00c0gata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features\nfor discriminative localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern\nRecognition (CVPR), 2016.\n\n10\n\n\f[23] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, Lubomir D. Bourdev, Ross B. Girshick, James Hays,\nPietro Perona, Deva Ramanan, Piotr Doll\u00e1r, and C. Lawrence Zitnick. Microsoft coco: Common objects in\ncontext. In Proceedings of the European Conference on Computer Vision (ECCV), 2014.\n\n[24] Matthew D. Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In Proceedings\n\nof the European Conference on Computer Vision (ECCV), 2014.\n\n[25] Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin A. Riedmiller. Striving for\nsimplicity: The all convolutional net. In International Conference on Learning Representations (ICLR),\n2014.\n\n[26] Sebastian Bach, Alexander Binder, Gr\u00e9goire Montavon, Frederick Klauschen, Klaus-Robert M\u00fcller,\nWojciech Samek, and Oscar Deniz Suarez. On pixel-wise explanations for non-linear classi\ufb01er decisions\nby layer-wise relevance propagation. In PloS one, 2015.\n\n[27] Jianming Zhang, Sarah Adel Bargal, Zhe L. Lin, Jonathan Brandt, Xiaohui Shen, and Stan Sclaroff.\nTop-down neural attention by excitation backprop. International Journal of Computer Vision (IJCV), 2017.\n[28] Piotr Dabkowski and Yarin Gal. Real time image saliency for black box classi\ufb01ers. In Advances in Neural\n\nInformation Processing Systems (NeurIPS), 2017.\n\n[29] Maruan Al-Shedivat, Kumar Avinava Dubey, and Eric P. Xing. Contextual explanation networks.\n\narXiv:1705.10301, 2018.\n\n[30] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing\nnetwork. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR),\npages 6230\u20136239, 2017.\n\n[31] Tobias Pohlen, Alexander Hermans, Markus Mathias, and Bastian Leibe. Full-resolution residual networks\nfor semantic segmentation in street scenes. In Proceedings of the IEEE Conference on Computer Vision\nand Pattern Recognition (CVPR), pages 3309\u20133318, 2017.\n\n[32] Guosheng Lin, Anton Milan, Chunhua Shen, and Ian D. Reid. Re\ufb01nenet: Multi-path re\ufb01nement networks\nfor high-resolution semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision\nand Pattern Recognition (CVPR), pages 5168\u20135177, 2017.\n\n[33] Evan Shelhamer, Jonathan Long, and Trevor Darrell. Fully convolutional networks for semantic segmen-\ntation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR),\n2015.\n\n[34] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan Loddon Yuille.\nSemantic image segmentation with deep convolutional nets and fully connected crfs. In International\nConference on Learning Representations (ICLR), 2015.\n\n[35] Fisher Yu and Vladlen Koltun. Multi-scale context aggregation by dilated convolutions. In International\n\nConference on Learning Representations (ICLR), 2016.\n\n[36] Guosheng Lin, Chunhua Shen, Ian D. Reid, and Anton van den Hengel. Ef\ufb01cient piecewise training of\ndeep structured models for semantic segmentation. In Proceedings of the IEEE Conference on Computer\nVision and Pattern Recognition (CVPR), 2016.\n\n[37] Shuai Zheng, Sadeep Jayasumana, Bernardino Romera-Paredes, Vibhav Vineet, Zhizhong Su, Dalong\nDu, Chang Huang, and Philip H. S. Torr. Conditional random \ufb01elds as recurrent neural networks. In\nProceedings of the IEEE International Conference on Computer Vision (ICCV), 2015.\n\n[38] Siddhartha Chandra and Iasonas Kokkinos. Fast, exact and multi-scale inference for semantic image\nsegmentation with deep gaussian crfs. In Proceedings of the European Conference on Computer Vision\n(ECCV), 2016.\n\n[39] Varun Jampani, Martin Kiefel, and Peter V. Gehler. Learning sparse high dimensional \ufb01lters: Image\n\ufb01ltering, dense crfs and bilateral neural networks. In Proceedings of the IEEE Conference on Computer\nVision and Pattern Recognition (CVPR), 2016.\n\n[40] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan Loddon Yuille.\nDeeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully\nconnected crfs. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 40:834\u2013848,\n2018.\n\n[41] Golnaz Ghiasi and Charless C. Fowlkes. Laplacian pyramid reconstruction and re\ufb01nement for semantic\n\nsegmentation. In Proceedings of the European Conference on Computer Vision (ECCV), 2016.\n\n[42] Seunghoon Hong, Hyeonwoo Noh, and Bohyung Han. Decoupled deep neural network for semi-supervised\n\nsemantic segmentation. In Advances in Neural Information Processing Systems (NeurIPS), 2015.\n\n[43] Hyeonwoo Noh, Seunghoon Hong, and Bohyung Han. Learning deconvolution network for semantic\nsegmentation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015.\n\n11\n\n\f[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical\n\nimage segmentation. In MICCAI, 2015.\n\n[45] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder\narchitecture for image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence\n(PAMI), 39:2481\u20132495, 2016.\n\n[46] Rudra P. K. Poudel, Stephan Liwicki, and Roberto Cipolla. Fast-scnn: Fast semantic segmentation network.\n\narXiv:1902.04502, 2019.\n\n[47] Adam Paszke, Abhishek Chaurasia, Sangpil Kim, and Eugenio Culurciello. Enet: A deep neural network\n\narchitecture for real-time semantic segmentation. arXiv:1606.02147, 2017.\n\n[48] Davide Mazzini. Guided upsampling network for real-time semantic segmentation. 2018.\n[49] Hengshuang Zhao, Xiaojuan Qi, Xiaoyong Shen, Jianping Shi, and Jiaya Jia. Icnet for real-time semantic\nsegmentation on high-resolution images. In Proceedings of the European Conference on Computer Vision\n(ECCV), 2018.\n\n[50] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition.\nIn Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages\n770\u2013778, 2016.\n\n[51] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recogni-\n\ntion. In International Conference on Learning Representations (ICLR), 2015.\n\n[52] SiteOrigin. Background image generator. http://bg.siteorigin.com/.\n[53] Subtle patterns. https://www.toptal.com/designers/subtlepatterns/.\n[54] Mark Lee and Zico Kolter. On physical adversarial patches for object detection. arXiv preprint\n\narXiv:1906.11897, 2019.\n\n12\n\n\f", "award": [], "sourceid": 3479, "authors": [{"given_name": "Lukas", "family_name": "Hoyer", "institution": "Bosch Center for Artificial Intelligence"}, {"given_name": "Mauricio", "family_name": "Munoz", "institution": "Bosch Center for Artificial Intelligence"}, {"given_name": "Prateek", "family_name": "Katiyar", "institution": "Bosch Center for Artificial Intelligence"}, {"given_name": "Anna", "family_name": "Khoreva", "institution": "Bosch Center for Artificial Intelligence"}, {"given_name": "Volker", "family_name": "Fischer", "institution": "Robert Bosch GmbH, Bosch Center for Artificial Intelligence"}]}