ECCV 2020 论文列表

Autonomous vehicles operate in a dynamic environment, where the speed with which a vehicle can perceive and react impacts the safety and efficacy of the system. LiDAR provides a central and prominent sensory modality that informs many existing perceptual systems including object detection, segmentation, motion estimation and action recognition. The latency for perceptual system based on point cloud data can be dominated by the amount of time for a complete rotational scan (e.g. 100 ms). This built-in data capture latency is artificial, and based on treating the point cloud as a camera image in order to leverage camera-inspired architectures. However, unlike CCD camera sensors, most LiDAR point cloud data is natively a {\it streaming} data source in which laser reflections are sequentially recorded based on the rotational angle of the laser beam. In this work, we explore how to build an object detector that removes this artificial latency constraint, and instead operates on native streaming data in order to significantly reduce latency. This approach has the added benefit of reducing the peak computational burden on inference hardware by spreading the computation over the complete scan sequence. We demonstrate a family of streaming detection systems based on sequential modeling through a series of modifications to the traditional object detection meta-architecture. We highlight how this model may achieve competitive if not superior predictive performance with state-of-the-art, traditional non-streaming detection systems while achieving significant latency gains (e.g. 1/15'th - 1/3'rd of peak latency). Our results show that operating on LiDAR data in its native streaming formulation offers several advantages for self driving object detection -- advantages that we hope will be useful for any LiDAR perception system where minimizing latency is critical for safe and efficient operation."

Super-resolution and denoising are ill-posed yet fundamental image restoration tasks. In blind settings, the degradation kernel or the noise level are unknown. This makes restoration even more challenging, notably for learning-based methods, as they tend to overfit to the degradation seen during training. We present an analysis, in the frequency domain, of degradation-kernel overfitting in super-resolution and introduce a conditional learning perspective that extends to both super-resolution and denoising. Building on our formulation, we propose a stochastic frequency masking of images used in training to regularize the networks and address the overfitting problem. Our technique improves state-of-the-art methods on blind super-resolution with different synthetic kernels, real super-resolution, blind Gaussian denoising, and real-image denoising."

In this paper, we propose novel stochastic modeling of various components of a continuous sign language recognition (CSLR) system that is based on the transformer encoder and connectionist temporal classification (CTC). Most importantly, We model each sign gloss with multiple states, and the number of states is a categorical random variable that follows a learned probability distribution, providing stochastic fine-grained labels for training the CTC decoder. We further propose a stochastic frame dropping mechanism and a gradient stopping method to deal with the severe overfitting problem in training the transformer model with CTC loss. These two methods also help reduce the training computation, both in terms of time and space, significantly. We evaluated our model on popular CSLR datasets, and show its effectiveness compared to the state-of-the-art methods."

Current bundle adjustment solvers such as the Levenberg-Marquardt (LM) algorithm are limited by the bottleneck in solving the Reduced Camera System (RCS) whose dimension is proportional to the camera number. When the problem is scaled up, this step is neither efficient in computation nor manageable for a single compute node. In this work, we propose a stochastic bundle adjustment algorithm which seeks to decompose the RCS approximately inside the LM iterations to improve the efficiency and scalability. It first reformulates the quadratic programming problem of an LM iteration based on the clustering of the visibility graph by introducing the equality constraints across clusters. Then, we propose to relax it into a chance constrained problem and solve it through sampled convex program. The relaxation is intended to eliminate the interdependence between clusters embodied by the constraints, so that a large RCS can be decomposed into independent linear sub-problems. Numerical experiments on unordered Internet image sets and sequential SLAM image sets, as well as distributed experiments on large-scale datasets, have demonstrated the high efficiency and scalability of the proposed approach."

Existing Particle Imaging Velocimetry techniques require the use of high-speed cameras to reconstruct time-resolved fluid flows. These cameras provides high-resolution images at high frame rates, which generates bandwidth and memory issues. By capturing only changes in the brightness with a very low latency and at low data rate, event-based cameras have the ability to tackle such issues. In this paper, we present a new framework that retrieves dense 3D measurements of the fluid velocity field using a pair of event-based cameras. First, we track particles inside the two event sequences in order to estimate their 2D velocity in the two sequences of images. A stereo-matching step is then performed to retrieve their 3D positions. These intermediate outputs are incorporated into an optimization framework that also includes physically plausible regularizers, in order to retrieve the 3D velocity field. Extensive experiments on both simulated and real data demonstrate the efficacy of our approach."

Recently, stacked networks show powerful performance in Image Restoration, such as challenging motion deblurring problems. However, the number of stacking levels is a hyper-parameter fine-tuned manually, making the stacking levels static during training without theoretical explanations for optimal settings. To address this challenge, we leverage the iterative process of the traditional plug-and-play method to provide a dynamic stacked network for Image Restoration. Specifically, a new degradation model with a novel update scheme is designed to integrate the deep neural network as the prior within the plug-and-play model. Compared with static stacked networks, our models are stacked dynamically during training via iterations, guided by a solid mathematical explanation. Theoretical proof on the convergence of the dynamic stacking process is provided. Experiments on the noise dataset BSD68, Set12, and motion blur dataset GoPro demonstrate that our framework outperforms the state-of-the-art in terms of PSNR and SSIM score without extra training process."

Most state-of-the-art deep neural networks are overparameterized and exhibit a high computational cost. A straightforward approach to this problem is to replace convolutional kernels with its low-rank tensor approximations, whereas the Canonical Polyadic tensor Decomposition is one of the most suited models. However, fitting the convolutional tensors by numerical optimization algorithms often encounters diverging components, i.e., extremely large rank-one tensors but canceling each other. Such degeneracy often causes the non-interpretable result and numerical instability for the neural network fine-tuning. This paper is the first study on degeneracy in the tensor decomposition of convolutional kernels. We present a novel method, which can stabilize the low-rank approximation of convolutional kernels and ensure efficient compression while preserving the high-quality performance of the neural networks. We evaluate our approach on popular CNN architectures for image classification and show that our method results in much lower accuracy degradation and provides consistent performance."

LiDAR point-cloud segmentation is an important problem for many applications. For large-scale point cloud segmentation, the extit{de facto} method is to project a 3D point cloud to get a 2D LiDAR image and use convolutions to process it. Despite the similarity between regular RGB and LiDAR images, we are the first to discover that the feature distribution of LiDAR images changes drastically at different image locations. Using standard convolutions to process such LiDAR images is problematic, as convolution filters pick up local features that are only active in specific regions in the image. As a result, the capacity of the network is under-utilized and the segmentation performance decreases. To fix this, we propose Spatially-Adaptive Convolution (SAC) to adopt different filters for different locations according to the input image. SAC can be computed efficiently since it can be implemented as a series of element-wise multiplications, im2col, and standard convolution. It is a general framework such that several previous methods can be seen as special cases of SAC. Using SAC, we build SqueezeSegV3 for LiDAR point-cloud segmentation and outperform all previous published methods by at least 2.0\% mIoU on the SemanticKITTI benchmark. Code and pretrained model are avalibale at \url{https://github.com/chenfengxu714/SqueezeSegV3}."

We propose the Square Attack, a score-based black-box $l_2$- and $l_\infty$- adversarial attack that does not rely on local gradient information and thus is not affected by gradient masking. Square Attack is based on a randomized search scheme which selects localized square-shaped updates at random positions so that at each iteration the perturbation is situated approximately at the boundary of the feasible set. Our method is significantly more query efficient and achieves a higher success rate compared to the state-of-the-art methods, especially in the untargeted setting. In particular, on ImageNet we improve the average query efficiency in the untargeted setting for various deep networks by a factor of at least $1.8$ and up to $3$ compared to the recent state-of-the-art $l_\infty$-attack of Al-Dujaili & O’Reilly [2]. Moreover, although our attack is black-box, it can also outperform gradient-based white-box attacks on the standard benchmarks achieving a new state-of-the-art in terms of the success rate."

In this paper we focus on the task of weakly-supervised semantic segmentation supervised with image-level labels. Since the pixel-level annotation is not available in the training process, we rely on region mining models to estimate the pseudo-masks from the image-level labels. Thus, in order to improve the final segmentation results, we aim to train a region-mining model which could accurately and completely highlight the target object regions for generating high-quality pseudo-masks. However, the region mining models are likely to only highlight the most discriminative regions instead of the entire objects. In this paper, we aim to tackle this problem from a novel perspective of optimization process. We propose a Splitting vs. Merging optimization strategy, which is mainly composed of the Discrepancy loss and the Intersection loss. The proposed Discrepancy loss aims at mining out regions of different spatial patterns instead of only the most discriminative region, which leads to the splitting effect. The Intersection loss aims at mining the common regions of the different maps, which leads to the merging effect. Our Splitting vs. Merging strategy helps to expand the output heatmap of the region mining model to the object scale. Finally, by training the segmentation model with the masks generated by our Splitting vs Merging strategy, we achieve the state-of-the-art weakly-supervised segmentation results on the Pascal VOC 2012 benchmark."

In this paper, motivated by human natural ability to perceive unseen surroundings imaginatively, we propose a novel Spiral Generative Network, SpiralNet, to perform image extrapolation in a spiral manner, which regards extrapolation as an evolution process growing from an input sub-image along a spiral curve to an expanded full image. Our SpiralNet, consisting of ImagineGAN and SliceGAN, disentangles image extrapolation problem into two independent sub-tasks as semantic structure prediction (via ImagineGAN) and contextual detail generation (via SliceGAN), making the whole task more tractable. The design of SliceGAN implicitly harnesses the correlation between generated contents and extrapolating direction, divide-and-conquer while generation-by-parts. Extensive experiments on datasets covering both objects and scenes under different cases show that our method achieves state-of-the-art performance on image extrapolation. We also conduct ablation study to validate efficacy of our design. Our code is available at https://github.com/zhenglab/spiralnet ."

Event-based cameras display great potential for a variety of tasks such as high-speed motion detection and navigation in low-light environments where conventional frame-based cameras suffer critically. This is attributed to their high temporal resolution, high dynamic range, and low-power consumption. However, conventional computer vision methods as well as deep Analog Neural Networks (ANNs) are not suited to work well with the asynchronous and discrete nature of event camera outputs. Spiking Neural Networks (SNNs) serve as ideal paradigms to handle event camera outputs, but deep SNNs suffer in terms of performance due to the spike vanishing phenomenon. To overcome these issues, we present Spike-FlowNet, a deep hybrid neural network architecture integrating SNNs and ANNs for efficiently estimating optical flow from sparse event camera outputs without sacrificing the performance. The network is end-to-end trained with self-supervised learning on Multi-Vehicle Stereo Event Camera (MVSEC) dataset. Spike-FlowNet outperforms its corresponding ANN-based method in terms of the optical flow prediction capability while providing significant computational efficiency."

Data augmentation in feature space is effective to increase data diversity. Previous methods assume that different classes have the same covariance in their feature distributions. Thus, feature transform between different classes is performed via translation. However, this approach is no longer valid for recent deep metric learning scenarios, where feature normalization is widely adopted and all features lie on a hypersphere. This work proposes a novel spherical feature transform approach. It relaxes the assumption of identical covariance between classes to an assumption of similar covariances of different classes on a hypersphere. Consequently, the feature transform is performed by a rotation that respects the spherical data distributions. We provide a simple and effective training method, and in depth analysis on the relation between the two different transforms. Comprehensive experiments on various deep metric learning benchmarks and different baselines verify that our method achieves consistent performance improvement and state-of-the-art results."

Speech-driven facial animation methods should produce accurate and realistic lip motions with natural expressions and realistic texture portraying target-specific facial characteristics. Moreover, the methods should also be adaptable to any unknown faces and speech quickly during inference. Current state-of-the-art methods fail to generate realistic animation from any speech on unknown faces due to their poor gen-eralization over different facial characteristics, languages, and accents. Some of these failures can be attributed to the end-to-end learning of the complex relationship between the multiple modalities of speech and the video. In this paper, we propose a novel strategy where we partition the problem and learn the motion and texture separately. Firstly, we train a GAN network to learn the lip motion in a canonical landmark using DeepSpeech features and induce eye-blinks before transferring the motion to the person-specific face. Next, we use another GAN based texture generator network to generate high fidelity face corresponding to the motion on person-specific landmark. We use meta-learning to make the texture generator GAN more flexible to adapt to the unknown subject’s traits of the face during inference. Our method gives significantly improved facial animation than the state-of-the-art methods and generalizes well across the different datasets, different languages, and accents,and also works reliably well in presence of noises in the speech."

Adversarial attacks are valuable for providing insights into the blind-spots of deep learning models and help improve their robustness. Existing work on adversarial attacks have mainly focused on static scenes; however, it remains unclear whether such attacks are effective against embodied agents, which could navigate and interact with a dynamic environment. In this work, we take the first step to study adversarial attacks for embodied agents. In particular, we generate spatiotemporal perturbations to form 3D adversarial examples, which exploit the interaction history in both the temporal and spatial dimensions. Regarding the temporal dimension, since agents make predictions based on historical observations, we develop a trajectory attention module to explore scene view contributions, which further help localize 3D objects appeared with highest stimuli. By conciliating with clues from the temporal dimension, along the spatial dimension, we adversarially perturb the physical properties (e.g., texture and 3D shape) of the contextual objects that appeared in the most important scene views. Extensive experiments on the EQA-v1 dataset for several embodied tasks in both the white-box and black-box settings have been conducted, which demonstrate that our perturbations have strong attack and generalization abilities."

Understanding crowd motion dynamics is critical to real-world applications, e.g., surveillance systems and autonomous driving. This is challenging because it requires effectively modeling the socially aware crowd spatial interaction and complex temporal dependencies. We believe attention is the most important factor for trajectory prediction. In this paper, we present STAR, a Spatio-Temporal grAph tRansformer framework, which tackles trajectory prediction by only attention mechanisms. STAR models intra-graph crowd interaction by TGConv, a novel Transformer-based graph convolution mechanism. The inter-graph temporal dependencies are modeled by separate temporal Transformers. STAR captures complex spatio-temporal interactions by interleaving between spatial and temporal Transformers. To calibrate the temporal prediction for the long-lasting effect of disappeared pedestrians, we introduce a read-writable external memory module, consistently being updated by the temporal Transformer. We show STAR outperforms the state-of-the-art models on 4 out of 5 real-world pedestrian trajectory prediction datasets and achieves comparable performance on the rest."

Textual cues are essential for everyday tasks like buying groceries and using public transport. To develop this assistive technology, we study the TextVQA task, i.e., reasoning about text in images to answer a question. Existing approaches are limited in their use of spatial relations and rely on fully-connected transformer-like architectures to implicitly learn the spatial structure of a scene. In contrast, we propose a novel spatially aware self-attention layer such that each visual entity only looks at neighboring entities defined by a spatial graph. Further, each head in our multi-head self-attention layer focuses on a different subset of relations. Our approach has two advantages: (1) each head considers local context instead of dispersing the attention amongst all visual entities; (2) we avoid learning redundant features. We show that our model improves the absolute accuracy of current state-of-the-art methods on TextVQA by 2.2% overall over an improved baseline, and 4.62% on questions that involve spatial reasoning and can be answered correctly using OCR tokens. Similarly on ST-VQA, we improve the absolute accuracy by 4.2%. We further show that spatially aware self-attention improves visual grounding."

In the feature maps of CNNs, there commonly exists considerable spatial redundancy that leads to much repetitive processing. Towards reducing this superfluous computation, we propose to compute features only at sparsely sampled locations, which are probabilistically chosen according to activation responses, and then densely reconstruct the feature map with an efficient interpolation procedure. With this sampling-interpolation scheme, our network avoids expending computation on spatial locations that can be effectively interpolated, while being robust to activation prediction errors through broadly distributed sampling. A technical challenge of this sampling-based approach is that the binary decision variables for representing discrete sampling locations are non-differentiable, making them incompatible with backpropagation. To circumvent this issue, we make use of a reparameterization trick based on the Gumbel-Softmax distribution, with which backpropagation can iterate these variables towards binary values. The presented network is experimentally shown to save substantial computation while maintaining accuracy over a variety of computer vision tasks."

Light field (LF) cameras record both intensity and directions of light rays, and capture scenes from a number of viewpoints. Both information within each perspective (i.e., spatial information) and among different perspectives (i.e., angular information) is beneficial to image super-resolution (SR). In this paper, we propose a spatial-angular interactive network (namely, LF-InterNet) for LF image SR. Specifically, spatial and angular features are first separately extracted from input LFs, and then repetitively interacted to progressively incorporate spatial and angular information. Finally, the interacted features are fused to superresolve each sub-aperture image. Experimental results demonstrate the superiority of LF-InterNet over the state-of-the-art methods, i.e., our method can achieve high PSNR and SSIM scores with low computational cost, and recover faithful details in the reconstructed images."