ECCV 2020 Papers

We introduce OmniSource, a novel framework for leveraging web data to train video recognition models. OmniSource overcomes the barriers between data formats, such as images, short videos, and long untrimmed videos for webly-supervised learning. First, data samples with multiple formats, curated by task-specific data collection and automatically filtered by a teacher model, are transformed into a unified form. Then a joint-training strategy is proposed to deal with the domain gaps between multiple data sources and formats in webly-supervised learning. Several good practices, including data balancing, resampling, and cross-dataset mixup are adopted in joint training. Experiments show that by utilizing data from multiple sources and formats, OmniSource is more data-efficient in training. With only 3.5M images and 800K minutes videos crawled from the internet without human labeling (less than 2% of prior works), our models learned with OmniSource improve Top-1 accuracy of 2D- and 3D-ConvNet baseline models by 3.0% and 3.9%, respectively, on the Kinetics-400 benchmark. With OmniSource, we establish new records with different pretraining strategies for video recognition. Our best models achieve 80.4%, 80.5%, and 83.6% Top-1 accuracies on the Kinetics-400 benchmark respectively for training-from-scratch, ImageNet pre-training and IG-65M pre-training."

In this paper, we introduce a new reinforcement learning (RL) based neural architecture search (NAS) methodology for effective and efficient generative adversarial network (GAN) architecture search. The key idea is to formulate the GAN architecture search problem as a Markov decision process (MDP) for smoother architecture sampling, which enables a more effective RL-based search algorithm by targeting the potential global optimal architecture. To improve efficiency, we exploit an off-policy GAN architecture search algorithm that makes efficient use of the samples generated by previous policies. Evaluation on two standard benchmark datasets (i.e., CIFAR-10 and STL-10) demonstrates that the proposed method is able to discover highly competitive architectures for generally better image generation results with a considerably reduced computational burden: 7 GPU hours. Our code is available at https://github.com/Yuantian013/E2GAN."

Anchor-based Siamese trackers have achieved remarkable advancements in accuracy, yet the further improvement is restricted by the lagged tracking robustness. We find the underlying reason is that the regression network in anchor-based methods is only trained on the positive anchor boxes ($IoU \geq 0.6$). This mechanism makes it difficult to refine the anchors whose overlap with the target objects are small. In this paper, we propose a novel object-aware anchor-free network to address this issue. First, instead of refining the reference anchor boxes, we directly predict the position and scale of target objects in an anchor-free fashion. Since each pixel in the groundtruth box is well trained, the tracker is capable of rectifying the weak predictions of target objects during inference. Second, we introduce a feature alignment module to learn an object-aware feature which corresponds with the predicted bounding box. The object-aware feature can further contribute to the classification of target object and background. Moreover, we present a novel Siamese tracking framework based on the anchor-free model. The experiments show that our anchor-free tracker achieves state-of-the-art performance on five benchmarks, including VOT-2018, VOT-2019, OTB-100, GOT-10k and LaSOT. The source code is available at https://github.com/researchmm/TracKit."

State-of-the-art navigation methods leverage a spatial memory to generalize to new environments, but their occupancy maps are limited to capturing the geometric structures directly observed by the agent. We propose occupancy anticipation, where the agent uses its egocentric RGB-D observations to infer the occupancy state beyond the visible regions. In doing so, the agent builds its spatial awareness more rapidly, which facilitates efficient exploration and navigation in 3D environments. By exploiting context in both the egocentric views and top-down maps our model successfully anticipates a broader map of the environment, with performance significantly better than strong baselines. Furthermore, when deploying our model for the sequential decision-making tasks of exploration and navigation, we outperform state-of-the-art methods on the Gibson and Matterport3D datasets. Our approach is the winning entry in the 2020 Habitat PointNav Challenge. Project page: http://vision.cs.utexas.edu/projects/occupancy_anticipation"

Pose estimation usually suffers from varying degrees of performance degeneration owing to occlusion. To conquer this dilemma, we propose an occlusion-aware siamese network to improve the performance. Specifically, we introduce scheme of feature erasing and reconstruction. Firstly, we utilize attention mechanism to predict the occlusion-aware attention map which is explicitly supervised and clean the feature map which is contaminated by different types of occlusions. Nevertheless, the cleaning procedure not only removes the useless information but also erases some valuable details. To overcome the defects caused by the erasing operation, we perform feature reconstruction to recover the information destroyed by occlusion and details lost in cleaning procedure. To make reconstructed features more precise and informative, we adopt siamese network equipped with OT divergence to guide the features of occluded images towards those of the un-occluded images. Algorithm is validated on MPII, LSP and COCO benchmarks and we achieve promising results."

We present a new learning-based method for multi-frame depth estimation from a color video, which is a fundamental problem in scene understanding, robot navigation or handheld 3D reconstruction. While recent learning-based methods estimate depth at high accuracy, 3D point clouds exported from their depth maps often fail to preserve important geometric feature (e.g., corners, edges, planes) of man-made scenes. Widely-used pixel-wise depth errors do not specifically penalize inconsistency on these features. These inaccuracies are particularly severe when subsequent depth reconstructions are accumulated in an attempt to scan a full environment with man-made objects with this kind of features. Our depth estimation algorithm therefore introduces a Combined Normal Map (CNM) constraint, which is designed to better preserve high-curvature features and global planar regions. In order to further improve the depth estimation accuracy, we introduce a new occlusion-aware strategy that aggregates initial depth predictions from multiple adjacent views into one final depth map and occlusion probability map for the current reference view. With the CNM constraint and the occlusion-aware loss, our method outperforms the state-of-the-art in terms of depth estimation accuracy, and preserves essential geometric features of man-made indoor scenes much better than other algorithms."

We present a scheme for fast environment light estimation from the RGBD appearance of individual objects and their local image areas. Conventional inverse rendering is too computationally demanding for real-time applications, and the performance of purely learning-based techniques may be limited by the meager input data available from individual objects. To address these issues, we propose an approach that takes advantage of physical principles from inverse rendering to constrain the solution, while also utilizing neural networks to expedite the more computationally expensive portions of its processing, to increase robustness to noisy input data as well as to improve temporal and spatial stability. This results in a rendering-aware system that estimates the local illumination distribution at an object with high accuracy and in real time. With the estimated lighting, virtual objects can be rendered in AR scenarios with shading that is consistent to the real scene, leading to improved realism."

Vision-and-Language Navigation (VLN) is unique in that it requires turning relatively general natural-language instructions into robot agent actions, on the basis of visible environments. This requires to extract value from two very different types of natural-language information. The first is object description (e.g., `table', `door'), each presenting as a tip for the agent to determine the next action by finding the item visible in the environment, and the second is action specification (e.g., `go straight', `turn left') which allows the robot to directly predict the next movements without relying on visual perceptions. However, most existing methods pay few attention to distinguish these information from each other during instruction encoding and mix together the matching between textual object/action encoding and visual perception/orientation features of candidate viewpoints. In this paper, we propose an Object-and-Action Aware Model (OAAM) that processes these two different forms of natural language based instruction separately. This enables each process to match object-centered/action-centered instruction to their own counterpart visual perception/action orientation flexibly. However, one side-issue caused by above solution is that an object mentioned in instructions may be observed in the direction of two or more candidate viewpoints, thus the OAAM may not predict the viewpoint on the shortest path as the next action. To handle this problem, we design a simple but effective path loss to penalize trajectories deviating from the ground truth path. Experimental results demonstrate the effectiveness of the proposed model and path loss, and the superiority of their combination with a 50% SPL score on the R2R dataset and a 40% CLS score on the R4R dataset in unseen environments, outperforming the previous state-of-the-art."

In this paper, we address the semantic segmentation problem with a focus on the context aggregation strategy. Our motivation is that the label of a pixel is the category of the object that the pixel belongs to. We present a simple yet effective approach, object-contextual representations, characterizing a pixel by exploiting the representation of the corresponding object class. First, we learn object regions under the supervision of the ground-truth segmentation. Second, we compute the object region representation by aggregating the representations of the pixels lying in the object region. Last, % the representation similarity we compute the relation between each pixel and each object region, and augment the representation of each pixel with the object-contextual representation which is a weighted aggregation of all the object region representations according to their relations with the pixel. We empirically demonstrate that the proposed approach achieves competitive performance on various challenging semantic segmentation benchmarks: Cityscapes, ADE20K, LIP, PASCAL-Context, and COCO-Stuff. Cityscapes, ADE20K, LIP, PASCAL-Context and COCO-Stuff. Our submission ""HRNet + OCR + SegFix"" achieves the 1-st place on the Cityscapes leaderboard by the time of submission. "

Accurate 3D object detection in LiDAR based point clouds suffers from the challenges of data sparsity and irregularities. Existing methods strive to organize the points regularly, e.g. voxelize, pass them through a designed 2D/3D neural network, and then define object-level anchors that predict offsets of 3D bounding boxes using collective evidences from all the points on the objects of interest. Contrary to the state-of-the-art anchor-based methods, based on the very nature of data sparsity, we observe that even points on an individual object part are informative about semantic information of the object. We thus argue in this paper for an approach opposite to existing methods using object-level anchors. Inspired by compositional models, which represent an object as parts and their spatial relations, we propose to represent an object as composition of its interior non-empty voxels, termed hotspots, and the spatial relations of hotspots. This gives rise to the representation of Object as Hotspots (OHS). Based on OHS, we further propose an anchor-free detection head with a novel ground truth assignment strategy that deals with inter-object point-sparsity imbalance to prevent the network from biasing towards objects with more points. Experimental results show that our proposed method works remarkably well on objects with a small number of points. Notably, our approach ranked 1st on KITTI 3D Detection Benchmark for cyclist and pedestrian detection, and achieved state-of-the-art performance on NuScenes 3D Detection Benchmark."

We introduce an object tracking algorithm that predicts the future locations of the target object and assists the tracker to handle object occlusion. Given a few frames of an object that are extracted from a complete input sequence, we aim to predict the object’s location in the future frames. To facilitate the future prediction ability, we follow three key observations: 1) object motion trajectory is affected significantly by camera motion; 2) the past trajectory of an object can act as a salient cue to estimate the object motion in the spatial domain; 3) previous frames contain the surroundings and appearance of the target object, which is useful for predicting the target object’s future locations. We incorporate these three observations into our method that employs a multi-stream convolutional-LSTM network. By combining the heatmap scores from our tracker (that utilises appearance inference) and the locations of the target object from our trajectory inference, we predict the final target’s location in each frame. Comprehensive evaluations show that our method sets new state-of-the-art performance on a few commonly used tracking benchmarks."

Given multiple datasets with different label spaces, the goal of this work is to train a single object detector predicting over the union of all the label spaces. The practical benefits of such an object detector are obvious and significant---application-relevant categories can be picked and merged form arbitrary existing datasets. However, naive merging of datasets is not possible in this case, due to inconsistent object annotations. Consider an object category like faces that is annotated in one dataset, but is not annotated in another dataset, although the object itself appears in the later's images. Some categories, like face here, would thus be considered foreground in one dataset, but background in another. To address this challenge, we design a framework which works with such partial annotations, and we exploit a pseudo labeling approach that we adapt for our specific case. We propose loss functions that carefully integrate partial but correct annotations with complementary but noisy pseudo labels. Evaluation in the proposed novel setting requires full annotation on the test set. We collect the required annotations and define a new challenging experimental setup for this task based on existing public datasets. We show improved performances compared to competitive baselines and appropriate adaptations of existing work."

In this paper, we consider the task of one-shot object detection, which consists in detecting objects defined by a single demonstration. Differently from the standard object detection, the classes of objects used for training and testing do not overlap. We build the one-stage system that performs localization and recognition jointly. We use dense correlation matching of learned local features to find correspondences, a feed-forward geometric transformation model to align features and bilinear resampling of the correlation tensor to compute the detection score of the aligned features. All the components are differentiable, which allows end-to-end training. Experimental evaluation on several challenging domains (retail products, 3D objects, buildings and logos) shows that our method can detect unseen classes (e.g., toothpaste when trained on groceries) and outperforms several baselines by a significant margin. Our code is available online: https://github.com/aosokin/os2d"

Blind deblurring methods are sensitive to outliers, such as saturated pixels and non-Gaussian noise. Even a small amount of outliers can dramatically degrade the quality of the estimated blur kernel, because the outliers are not conforming to the linear formation of the blurring process. Prior arts develop sophisticated edge-selecting steps or noise filtering pre-processing steps to deal with outliers (i.e. indirect approaches). However, these indirect approaches may fail when massive outliers are presented, since informative details may be polluted by outliers or erased during the pre-processing steps. To address these problems, this paper develops a simple yet effective Outlier Identifying and Discarding (OID) method, which alleviates limitations in existing Maximum A Posteriori (MAP)- based deblurring models when significant outliers are presented. Unlike previous indirect outlier processing methods, OID tackles outliers directly by explicitly identifying and discarding them, when updating both the latent image and the blur kernel during the deblurring process, where the outliers are detected by using the sparse and entropy-based modules. OID is easy to implement and extendable for non-blind restoration. Extensive experiments demonstrate the superiority of OID against recent works both quantitatively and qualitatively."

We propose to learn invariant representations, in the data domain, to achieve interpretability in algorithmic fairness. Invariance implies a selectivity for high level, relevant correlations w.r.t. class label annotations, and a robustness to irrelevant correlations with protected characteristics such as race or gender. We introduce a non-trivial setup in which the training set exhibits a strong bias such that class label annotations are irrelevant and spurious correlations cannot be distinguished. To address this problem, we introduce an adversarially trained model with a null-sampling procedure to produce invariant representations in the data domain. To enable disentanglement, a partially-labelled representative set is used. By placing the representations into the data domain, the changes made by the model are easily examinable by human auditors. We show the effectiveness of our method on both image and tabular datasets: Coloured MNIST, the CelebA and the Adult dataset."

Novel view synthesis often requires to have the paired data from both the source and target views. This paper proposes a view translation model within cVAE-GAN framework for the purpose of unpaired training. We design a conditional deformable module (CDM) which uses the source (or target) view condition vectors as the filters to convolve the feature maps from the main branch. It generates several pairs of displacement maps like the 2D optical flows. The flows then deform the features, and the results are given to the main branch of the encoder (or decoder) by the deformed feature based normalization module (DFNM). DFNM scales and offsets the feature maps in the main branch given its deformed input from the side branch. With the CDM and DFNM, the encoder outputs a view-irrelevant posterior, while the decoder takes the sample drawn from it to synthesize the reconstructed and the view-translated images. To further ensure the disentanglement between the views and other factors, we add adversarial training on the code drawn from the view-irrelevant posterior. The results and the ablation study on MultiPIE and 3D chair datasets validate the effectiveness of the whole framework and the designed module."

but also Listen: Learning Multimodal Violence Detection under Weak Supervision","Violence detection has been studied in computer vision for years. However, previous work are either superficial, e.g., classification of short-clips, and the single scenario, or undersupplied, e.g., the single modality, and hand-crafted features based multimodality. To address this problem, in this work we first release a large-scale and multi-scene dataset named XD-Violence with a total duration of 217 hours, containing 4754 untrimmed videos with audio signals and weak labels. Then we propose a neural network containing three parallel branches to capture different relations among video snippets and integrate features, where holistic branch captures long-range dependencies using similarity prior, localized branch captures local positional relation using proximity prior, and score branch dynamically captures the closeness of predicted score. Besides, our method also includes an approximator to meet the needs of online detection. Our method outperforms other state-of-the-art methods on our released dataset and other existing benchmark. Moreover, extensive experimental results also show the positive effect of multimodal (audio-visual) input and modeling relationships. The code and dataset will be released in https://roc-ng.github.io/XD-Violence/."

We propose NormalGAN, a fast adversarial learning-based method to reconstruct the complete and detailed 3D human from a single RGB-D image. Given a single front-view RGB-D image, NormalGAN performs two steps: front-view RGB-D rectification and back-view RGB-D inference. The final model was then generated by simply combining the front-view and back-view RGB-D information. However, inferring back-view RGB-D image with high-quality geometric details and plausible texture is not trivial. Our key observation is: Normal maps generally encode much more information of 3D surface details than RGB and depth images. Therefore, learning geometric details from normal maps is superior than other representations. In NormalGAN, an adversarial learning framework conditioned by normal maps is introduced, which is used to not only improve the front-view depth denoising performance, but also infer back-view depth image with surprisingly geometric details. Moreover, for texture recovery, we remove shading information from the front-view RGB image based on the refined normal map, which further improves the quality of the back-view color inference. Results and experiments on both testing data set and real captured data demonstrate the superior performance of our approach. Given a consumer RGB-D sensor, NormalGAN can generate the complete and detailed 3D human reconstruction results in 20 fps, which further enables convenient interactive experiences in telepresence, AR/VR and gaming scenarios. "

In this paper, we propose a robust and efficient end-to-end non-local spatial propagation network for depth completion. The proposed network takes RGB and sparse depth images as inputs and estimates non-local neighbors and their affinities of each pixel, as well as an initial depth map with pixel-wise confidences. The initial depth prediction is then iteratively refined by its confidence and non-local spatial propagation procedure based on the predicted non-local neighbors and corresponding affinities. Unlike previous algorithms that utilize fixed-local neighbors, the proposed algorithm effectively avoids irrelevant local neighbors and concentrates on relevant non-local neighbors during propagation. In addition, we introduce a learnable affinity normalization to better learn the affinity combinations compared to conventional methods. The proposed algorithm is inherently robust to the mixed-depth problem on depth boundaries, which is one of the major issues for existing depth estimation/completion algorithms. Experimental results on indoor and outdoor datasets demonstrate that the proposed algorithm is superior to conventional algorithms in terms of depth completion accuracy and robustness to the mixed-depth problem."