CVPR 2020 论文列表

We consider the problem of space-time super-resolution (ST-SR): increasing spatial resolution of video frames and simultaneously interpolating frames to increase the frame rate. Modern approaches handle these axes one at a time. In contrast, our proposed model called STARnet super-resolves jointly in space and time. This allows us to leverage mutually informative relationships between time and space: higher resolution can provide more detailed information about motion, and higher frame-rate can provide better pixel alignment. The components of our model that generate latent low- and high-resolution representations during ST-SR can be used to finetune a specialized mechanism for just spatial or just temporal super-resolution. Experimental results demonstrate that STARnet improves the performances of space-time, spatial, and temporal video super-resolution by substantial margins on publicly available datasets.

Point clouds are useful in many applications like autonomous driving and robotics as they provide natural 3D information of the surrounding environments. While there are extensive research on 3D point clouds, scene understanding on 4D point clouds, a series of consecutive 3D point clouds frames, is an emerging topic and yet under-investigated. With 4D point clouds (3D point cloud videos), robotic systems could enhance their robustness by leveraging the temporal information from previous frames. However, the existing semantic segmentation methods on 4D point clouds suffer from low precision due to the spatial and temporal information loss in their network structures. In this paper, we propose SpSequenceNet to address this problem. The network is designed based on 3D sparse convolution. And we introduce two novel modules, a cross-frame global attention module and a cross-frame local interpolation module, to capture spatial and temporal information in 4D point clouds. We conduct extensive experiments on SemanticKITTI, and achieve the state-of-the-art result of 43.1% on mIoU, which is 1.5% higher than the previous best approach.

Human action is naturally compositional: humans can easily recognize and perform actions with objects that are different from those used in training demonstrations. In this paper, we study the compositionality of action by looking into the dynamics of subject-object interactions. We propose a novel model which can explicitly reason about the geometric relations between constituent objects and an agent performing an action. To train our model, we collect dense object box annotations on the Something-Something dataset. We propose a novel compositional action recognition task where the training combinations of verbs and nouns do not overlap with the test set. The novel aspects of our model are applicable to activities with prominent object interaction dynamics and to objects which can be tracked using state-of-the-art approaches; for activities without clearly defined spatial object-agent interactions, we rely on baseline scene-level spatio-temporal representations. We show the effectiveness of our approach not only on the proposed compositional action recognition task but also in a few-shot compositional setting which requires the model to generalize across both object appearance and action category.

ImageNet pre-training has long been considered crucial by the fine-grained sketch-based image retrieval (FG-SBIR) community due to the lack of large sketch-photo paired datasets for FG-SBIR training. In this paper, we propose a self-supervised alternative for representation pre-training. Specifically, we consider the jigsaw puzzle game of recomposing images from shuffled parts. We identify two key facets of jigsaw task design that are required for effective FG-SBIR pre-training. The first is formulating the puzzle in a mixed-modality fashion. Second we show that framing the optimisation as permutation matrix inference via Sinkhorn iterations is more effective than the common classifier formulation of Jigsaw self-supervision. Experiments show that this self-supervised pre-training strategy significantly outperforms the standard ImageNet-based pipeline across all four product-level FG-SBIR benchmarks. Interestingly it also leads to improved cross-category generalisation across both pre-train/fine-tune and fine-tune/testing stages.

Jigsaw puzzle solving is an intriguing problem which has been explored in computer vision for decades. This paper focuses on a specific variant of the problem--solving puzzles with eroded boundaries. Such erosion makes the problem extremely difficult, since most existing solvers utilize solely the information at the boundaries. Nevertheless, this variant is important since erosion and missing data often occur at the boundaries. The key idea of our proposed approach is to inpaint the eroded boundaries between puzzle pieces and later leverage the quality of the inpainted area to classify a pair of pieces as "neighbors or not". An interesting feature of our architecture is that the same GAN discriminator is used for both inpainting and classification; training of the second task is simply a continuation of the training of the first, beginning from the point it left off. We show that our approach outperforms other SOTA methods.

Differentiable image sampling in the form of backward warping has seen broad adoption in tasks like depth estimation and optical flow prediction. In contrast, how to perform forward warping has seen less attention, partly due to additional challenges such as resolving the conflict of mapping multiple pixels to the same target location in a differentiable way. We propose softmax splatting to address this paradigm shift and show its effectiveness on the application of frame interpolation. Specifically, given two input frames, we forward-warp the frames and their feature pyramid representations based on an optical flow estimate using softmax splatting. In doing so, the softmax splatting seamlessly handles cases where multiple source pixels map to the same target location. We then use a synthesis network to predict the interpolation result from the warped representations. Our softmax splatting allows us to not only interpolate frames at an arbitrary time but also to fine tune the feature pyramid and the optical flow. We show that our synthesis approach, empowered by softmax splatting, achieves new state-of-the-art results for video frame interpolation.

Better machine understanding of pedestrian behaviors enables faster progress in modeling interactions between agents such as autonomous vehicles and humans. Pedestrian trajectories are not only influenced by the pedestrian itself but also by interaction with surrounding objects. Previous methods modeled these interactions by using a variety of aggregation methods that integrate different learned pedestrians states. We propose the Social Spatio-Temporal Graph Convolutional Neural Network (Social-STGCNN), which substitutes the need of aggregation methods by modeling the interactions as a graph. Our results show an improvement over the state of art by 20% on the Final Displacement Error (FDE) and an improvement on the Average Displacement Error (ADE) with 8.5 times less parameters and up to 48 times faster inference speed than previously reported methods. In addition, our model is data efficient, and exceeds previous state of the art on the ADE metric with only 20% of the training data. We propose a kernel function to embed the social interactions between pedestrians within the adjacency matrix. Through qualitative analysis, we show that our model inherited social behaviors that can be expected between pedestrians trajectories. Code is available at https://github.com/abduallahmohamed/Social-STGCNN.

Most of the existing person re-identification (re-ID) methods achieve promising accuracy in a supervised manner, but they assume the identity labels of the target domain is available. This greatly limits the scalability of person re-ID in real-world scenarios. Therefore, the current person re-ID community focuses on the cross-domain person re-ID that aims to transfer the knowledge from a labeled source domain to an unlabeled target domain and exploits the specific knowledge from the data distribution of the target domain to further improve the performance. To reduce the gap between the source and target domains, we propose a Smoothing Adversarial Domain Attack (SADA) approach that guides the source domain images to align the target domain images by using a trained camera classifier. To stabilize a memory trace of cross-domain knowledge transfer after its initial acquisition from the source domain, we propose a p-Memory Reconsolidation (pMR) method that reconsolidates the source knowledge with a small probability p during the self-training of the target domain. With both SADA and pMR, the proposed method significantly improves the cross-domain person re-ID. Extensive experiments on Market-1501 and DukeMTMC-reID benchmarks show that our pMR-SADA outperforms all of the state-of-the-arts by a large margin.

Temporal convolution has been widely used for video classification. However, it is performed on spatio-temporal contexts in a limited view, which often weakens its capacity of learning video representation. To alleviate this problem, we propose a concise and novel SmallBig network, with the cooperation of small and big views. For the current time step, the small view branch is used to learn the core semantics, while the big view branch is used to capture the contextual semantics. Unlike traditional temporal convolution, the big view branch can provide the small view branch with the most activated video features from a broader 3D receptive field. Via aggregating such big-view contexts, the small view branch can learn more robust and discriminative spatio-temporal representations for video classification. Furthermore, we propose to share convolution in the small and big view branch, which improves model compactness as well as alleviates overfitting. As a result, our SmallBigNet achieves a comparable model size like 2D CNNs, while boosting accuracy like 3D CNNs. We conduct extensive experiments on the large-scale video benchmarks, e.g., Kinetics400, Something-Something V1 and V2. Our SmallBig network outperforms a number of recent state-of-the-art approaches, in terms of accuracy and/or efficiency. The codes and models will be available on https://github.com/xhl-video/SmallBigNet.

Sketchformer is a novel transformer-based representation for encoding free-hand sketches input in a vector form, i.e. as a sequence of strokes. Sketchformer effectively addresses multiple tasks: sketch classification, sketch based image retrieval (SBIR), and the reconstruction and interpolation of sketches. We report several variants exploring continuous and tokenized input representations, and contrast their performance. Our learned embedding, driven by a dictionary learning tokenization scheme, yields state of the art performance in classification and image retrieval tasks, when compared against baseline representations driven by LSTM sequence to sequence architectures: SketchRNN and derivatives. We show that sketch reconstruction and interpolation are improved significantly by the Sketchformer embedding for complex sketches with longer stroke sequences.

Previous researches of sketches often considered sketches in pixel format and leveraged CNN based models in the sketch understanding. Fundamentally, a sketch is stored as a sequence of data points, a vector format representation, rather than the photo-realistic image of pixels. SketchRNN studied a generative neural representation for sketches of vector format by Long Short Term Memory networks (LSTM). Unfortunately, the representation learned by SketchRNN is primarily for the generation tasks, rather than the other tasks of recognition and retrieval of sketches. To this end and inspired by the recent BERT model, we present a model of learning Sketch Bidirectional Encoder Representation from Transformer (Sketch-BERT). We generalize BERT to sketch domain, with the novel proposed components and pre-training algorithms, including the newly designed sketch embedding networks, and the self-supervised learning of sketch gestalt. Particularly, towards the pre-training task, we present a novel Sketch Gestalt Model (SGM) to help train the Sketch-BERT. Experimentally, we show that the learned representation of Sketch-BERT can help and improve the performance of the downstream tasks of sketch recognition, sketch retrieval, and sketch gestalt.

A recent strand of work in view synthesis uses deep learning to generate multiplane images--a camera-centric, layered 3D representation--given two or more input images at known viewpoints. We apply this representation to single-view view synthesis, a problem which is more challenging but has potentially much wider application. Our method learns to predict a multiplane image directly from a single image input, and we introduce scale-invariant view synthesis for supervision, enabling us to train on online video. We show this approach is applicable to several different datasets, that it additionally generates reasonable depth maps, and that it learns to fill in content behind the edges of foreground objects in background layers.

Deep learning models have shown impressive performance across a spectrum of computer vision applications including medical diagnosis and autonomous driving. One of the major concerns that these models face is their susceptibility to adversarial attacks. Realizing the importance of this issue, more researchers are working towards developing robust models that are less affected by adversarial attacks. Adversarial training method shows promising results in this direction. In adversarial training regime, models are trained with mini-batches augmented with adversarial samples. Fast and simple methods (e.g., single-step gradient ascent) are used for generating adversarial samples, in order to reduce computational complexity. It is shown that models trained using single-step adversarial training method (adversarial samples are generated using non-iterative method) are pseudo robust. Further, this pseudo robustness of models is attributed to the gradient masking effect. However, existing works fail to explain when and why gradient masking effect occurs during single-step adversarial training. In this work, (i) we show that models trained using single-step adversarial training method learn to prevent the generation of single-step adversaries, and this is due to over-fitting of the model during the initial stages of training, and (ii) to mitigate this effect, we propose a single-step adversarial training method with dropout scheduling. Unlike models trained using existing single-step adversarial training methods, models trained using the proposed single-step adversarial training method are robust against both single-step and multi-step adversarial attacks, and the performance is on par with models trained using computationally expensive multi-step adversarial training methods, in white-box and black-box settings.

Recent years have seen a rapid growth in new approaches improving the accuracy of semantic segmentation in a weakly supervised setting, i.e. with only image-level labels available for training. However, this has come at the cost of increased model complexity and sophisticated multi-stage training procedures. This is in contrast to earlier work that used only a single stage -- training one segmentation network on image labels -- which was abandoned due to inferior segmentation accuracy. In this work, we first define three desirable properties of a weakly supervised method: local consistency, semantic fidelity, and completeness. Using these properties as guidelines, we then develop a segmentation-based network model and a self-supervised training scheme to train for semantic masks from image-level annotations in a single stage. We show that despite its simplicity, our method achieves results that are competitive with significantly more complex pipelines, substantially outperforming earlier single-stage methods.

Most recent 6D pose estimation frameworks first rely on a deep network to establish correspondences between 3D object keypoints and 2D image locations and then use a variant of a RANSAC-based Perspective-n-Point (PnP) algorithm. This two-stage process, however, is suboptimal: First, it is not end-to-end trainable. Second, training the deep network relies on a surrogate loss that does not directly reflect the final 6D pose estimation task. In this work, we introduce a deep architecture that directly regresses 6D poses from correspondences. It takes as input a group of candidate correspondences for each 3D keypoint and accounts for the fact that the order of the correspondences within each group is irrelevant, while the order of the groups, that is, of the 3D keypoints, is fixed. Our architecture is generic and can thus be exploited in conjunction with existing correspondence-extraction networks so as to yield single-stage 6D pose estimation frameworks. Our experiments demonstrate that these single-stage frameworks consistently outperform their two-stage counterparts in terms of both accuracy and speed.

Existing domain generalization methods for face anti-spoofing endeavor to extract common differentiation features to improve the generalization. However, due to large distribution discrepancies among fake faces of different domains, it is difficult to seek a compact and generalized feature space for the fake faces. In this work, we propose an end-to-end single-side domain generalization framework (SSDG) to improve the generalization ability of face anti-spoofing. The main idea is to learn a generalized feature space, where the feature distribution of the real faces is compact while that of the fake ones is dispersed among domains but compact within each domain. Specifically, a feature generator is trained to make only the real faces from different domains undistinguishable, but not for the fake ones, thus forming a single-side adversarial learning. Moreover, an asymmetric triplet loss is designed to constrain the fake faces of different domains separated while the real ones aggregated. The above two points are integrated into a unified framework in an end-to-end training manner, resulting in a more generalized class boundary, especially good for samples from novel domains. Feature and weight normalization is incorporated to further improve the generalization ability. Extensive experiments show that our proposed approach is effective and outperforms the state-of-the-art methods on four public databases. The code is released online.

Recovering a high dynamic range (HDR) image from a single low dynamic range (LDR) input image is challenging due to missing details in under-/over-exposed regions caused by quantization and saturation of camera sensors. In contrast to existing learning-based methods, our core idea is to incorporate the domain knowledge of the LDR image formation pipeline into our model. We model the HDR-to-LDR image formation pipeline as the (1) dynamic range clipping, (2) non-linear mapping from a camera response function, and (3) quantization. We then propose to learn three specialized CNNs to reverse these steps. By decomposing the problem into specific sub-tasks, we impose effective physical constraints to facilitate the training of individual sub-networks. Finally, we jointly fine-tune the entire model end-to-end to reduce error accumulation. With extensive quantitative and qualitative experiments on diverse image datasets, we demonstrate that the proposed method performs favorably against state-of-the-art single-image HDR reconstruction algorithms.

We address the problem of removing undesirable reflections from a single image captured through a glass surface, which is an ill-posed, challenging but practically important problem for photo enhancement. Inspired by iterative structure reduction for hidden community detection in social networks, we propose an Iterative Boost Convolutional LSTM Network (IBCLN) that enables cascaded prediction for reflection removal. IBCLN is a cascaded network that iteratively refines the estimates of transmission and reflection layers in a manner that they can boost the prediction quality to each other, and information across steps of the cascade is transferred using an LSTM. The intuition is that the transmission is the strong, dominant structure while the reflection is the weak, hidden structure. They are complementary to each other in a single image and thus a better estimate and reduction on one side from the original image leads to a more accurate estimate on the other side. To facilitate training over multiple cascade steps, we employ LSTM to address the vanishing gradient problem, and propose residual reconstruction loss as further training guidance. Besides, we create a dataset of real-world images with reflection and ground-truth transmission layers to mitigate the problem of insufficient data. Comprehensive experiments demonstrate that the proposed method can effectively remove reflections in real and synthetic images compared with state-of-the-art reflection removal methods.

Event cameras are bio-inspired sensors that asynchronously report intensity changes in microsecond resolution. DAVIS can capture high dynamics of a scene and simultaneously output high temporal resolution events and low frame-rate intensity images. In this paper, we propose a single image (potentially blurred) and events based optical flow estimation approach. First, we demonstrate how events can be used to improve flow estimates. To this end, we encode the relation between flow and events effectively by presenting an event-based photometric consistency formulation. Then, we consider the special case of image blur caused by high dynamics in the visual environments and show that including the blur formation in our model further constrains flow estimation. This is in sharp contrast to existing works that ignore the blurred images while our formulation can naturally handle either blurred or sharp images to achieve accurate flow estimation. Finally, we reduce flow estimation, as well as image deblurring, to an alternative optimization problem of an objective function using the primal-dual algorithm. Experimental results on both synthetic and real data (with blurred and non-blurred images) show the superiority of our model in comparison to state-of-the-art approaches.