CVPR 2020 论文列表

Current popular online multi-object tracking (MOT) solutions apply single object trackers (SOTs) to capture object motions, while often requiring an extra affinity network to associate objects, especially for the occluded ones. This brings extra computational overhead due to repetitive feature extraction for SOT and affinity computation. Meanwhile, the model size of the sophisticated affinity network is usually non-trivial. In this paper, we propose a novel MOT framework that unifies object motion and affinity model into a single network, named UMA, in order to learn a compact feature that is discriminative for both object motion and affinity measure. In particular, UMA integrates single object tracking and metric learning into a unified triplet network by means of multi-task learning. Such design brings advantages of improved computation efficiency, low memory requirement and simplified training procedure. In addition, we equip our model with a task-specific attention module, which is used to boost task-aware feature learning. The proposed UMA can be easily trained end-to-end, and is elegant - requiring only one training stage. Experimental results show that it achieves promising performance on several MOT Challenge benchmarks.

Among the major remaining challenges for generative adversarial networks (GANs) is the capacity to synthesize globally and locally coherent images with object shapes and textures indistinguishable from real images. To target this issue we propose an alternative U-Net based discriminator architecture, borrowing the insights from the segmentation literature. The proposed U-Net based architecture allows to provide detailed per-pixel feedback to the generator while maintaining the global coherence of synthesized images, by providing the global image feedback as well. Empowered by the per-pixel response of the discriminator, we further propose a per-pixel consistency regularization technique based on the CutMix data augmentation, encouraging the U-Net discriminator to focus more on semantic and structural changes between real and fake images. This improves the U-Net discriminator training, further enhancing the quality of generated samples. The novel discriminator improves over the state of the art in terms of the standard distribution and image quality metrics, enabling the generator to synthesize images with varying structure, appearance and levels of detail, maintaining global and local realism. Compared to the BigGAN baseline, we achieve an average improvement of 2.7 FID points across FFHQ, CelebA, and the proposed COCO-Animals dataset.

Semi-supervised video object segmentation aims to separate a target object from a video sequence, given the mask in the first frame. Most of current prevailing methods utilize information from additional modules trained in other domains like optical flow and instance segmentation, and as a result they do not compete with other methods on common ground. To address this issue, we propose a simple yet strong transductive method, in which additional modules, datasets, and dedicated architectural designs are not needed. Our method takes a label propagation approach where pixel labels are passed forward based on feature similarity in an embedding space. Different from other propagation methods, ours diffuses temporal information in a holistic manner which take accounts of long-term object appearance. In addition, our method requires few additional computational overhead, and runs at a fast 37 fps speed. Our single model with a vanilla ResNet50 backbone achieves an overall score of 72.3% on the DAVIS 2017 validation set and 63.1% on the test set. This simple yet high performing and efficient method can serve as a solid baseline that facilitates future research. Code and models are available at https://github.com/ microsoft/transductive-vos.pytorch.

Human motion prediction, the task of predicting future 3D human poses given a sequence of observed ones, has been mostly treated as a deterministic problem. However, human motion is a stochastic process: Given an observed sequence of poses, multiple future motions are plausible. Existing approaches to modeling this stochasticity typically combine a random noise vector with information about the previous poses. This combination, however, is done in a deterministic manner, which gives the network the flexibility to learn to ignore the random noise. Alternatively, in this paper, we propose to stochastically combine the root of variations with previous pose information, so as to force the model to take the noise into account. We exploit this idea for motion prediction by incorporating it into a recurrent encoder-decoder network with a conditional variational autoencoder block that learns to exploit the perturbations. Our experiments on two large-scale motion prediction datasets demonstrate that our model yields high-quality pose sequences that are much more diverse than those from state-of-the-art stochastic motion prediction techniques.

Dynamic medical images are often limited in its application due to the large radiation doses and longer image scanning and reconstruction times. Existing methods attempt to reduce the volume samples in the dynamic sequence by interpolating the volumes between the acquired samples. However, these methods are limited to either 2D images and/or are unable to support large but periodic variations in the functional motion between the image volume samples. In this paper, we present a spatiotemporal volumetric interpolation network (SVIN) designed for 4D dynamic medical images. SVIN introduces dual networks: the first is the spatiotemporal motion network that leverages the 3D convolutional neural network (CNN) for unsupervised parametric volumetric registration to derive spatiotemporal motion field from a pair of image volumes; the second is the sequential volumetric interpolation network, which uses the derived motion field to interpolate image volumes, together with a new regression-based module to characterize the periodic motion cycles in functional organ structures. We also introduce an adaptive multi-scale architecture to capture the volumetric large anatomy motions. Experimental results demonstrated that our SVIN outperformed state-of-the-art temporal medical interpolation methods and natural video interpolation method that has been extended to support volumetric images. Code is available at [1].

Recently, deep learning has been explored as a promising direction for image compression. Removing the spatial redundancy of the image is crucial for image compression and most learning based methods focus on removing the redundancy between adjacent pixels. Intuitively, to explore larger pixel range beyond adjacent pixel is beneficial for removing the redundancy. In this paper, we propose a fast yet effective method for end-to-end image compression by incorporating a novel spatial recurrent neural network. Block based LSTM is utilized to remove the redundant information between adjacent pixels and blocks. Besides, the proposed method is a potential efficient system that parallel computation on individual blocks is possible. Experimental results demonstrate that the proposed model outperforms state-of-the-art traditional image compression standards and learning based image compression models in terms of both PSNR and MS-SSIM metrics. It provides a 26.73% bits-reduction than High Efficiency Video Coding (HEVC), which is the current official state-of-the-art video codec.

Many computer vision applications require robust and efficient estimation of camera geometry. The robust estimation is usually based on solving camera geometry problems from a minimal number of input data measurements, i.e. solving minimal problems in a RANSAC framework. Minimal problems often result in complex systems of polynomial equations. Many state-of-the-art efficient polynomial solvers to these problems are based on Grobner basis and the action-matrix method that has been automatized and highly optimized in recent years. In this paper we study an alternative algebraic method for solving systems of polynomial equations, i.e., the sparse resultant-based method and propose a novel approach to convert the resultant constraint to an eigenvalue problem. This technique can significantly improve the efficiency and stability of existing resultant-based solvers. We applied our new resultant-based method to a large variety of computer vision problems and show that for most of the considered problems, the new method leads to solvers that are the same size as the the best available Grobner basis solvers and of similar accuracy. For some problems the new sparse-resultant based method leads to even smaller and more stable solvers than the state-of-the-art Grobner basis solvers. Our new method can be fully automatized and incorporated into existing tools for automatic generation of efficient polynomial solvers and as such it represents a competitive alternative to popular Grobner basis methods for minimal problems in computer vision.

In this work, we develop a shared multi-attention model for multi-label zero-shot learning. We argue that designing attention mechanism for recognizing multiple seen and unseen labels in an image is a non-trivial task as there is no training signal to localize unseen labels and an image only contains a few present labels that need attentions out of thousands of possible labels. Therefore, instead of generating attentions for unseen labels which have unknown behaviors and could focus on irrelevant regions due to the lack of any training sample, we let the unseen labels select among a set of shared attentions which are trained to be label-agnostic and to focus on only relevant/foreground regions through our novel loss. Finally, we learn a compatibility function to distinguish labels based on the selected attention. We further propose a novel loss function that consists of three components guiding the attention to focus on diverse and relevant image regions while utilizing all attention features. By extensive experiments, we show that our method improves the state of the art by 2.9% and 1.4% F1 score on the NUS-WIDE and the large scale Open Images datasets, respectively.

Neural architecture search (NAS) aims to automatically design deep neural networks of satisfactory performance. Wherein, architecture performance predictor is critical to efficiently value an intermediate neural architecture. But for the training of this predictor, a number of neural architectures and their corresponding real performance often have to be collected. In contrast with classical performance predictor optimized in a fully supervised way, this paper suggests a semi-supervised assessor of neural architectures. We employ an auto-encoder to discover meaningful representations of neural architectures. Taking each neural architecture as an individual instance in the search space, we construct a graph to capture their intrinsic similarities, where both labeled and unlabeled architectures are involved. A graph convolutional neural network is introduced to predict the performance of architectures based on the learned representations and their relation modeled by the graph. Extensive experimental results on the NAS-Benchmark-101 dataset demonstrated that our method is able to make a significant reduction on the required fully trained architectures for finding efficient architectures.

Adversarial examples can cause catastrophic mistakes in Deep Neural Network (DNNs) based vision systems e.g., for classification, segmentation and object detection. The vulnerability of DNNs against such attacks can prove a major roadblock towards their real-world deployment. Transferability of adversarial examples demand generalizable defenses that can provide cross-task protection. Adversarial training that enhances robustness by modifying target model's parameters lacks such generalizability. On the other hand, different input processing based defenses fall short in the face of continuously evolving attacks. In this paper, we take the first step to combine the benefits of both approaches and propose a self-supervised adversarial training mechanism in the input space. By design, our defense is a generalizable approach and provides significant robustness against the unseen adversarial attacks (e.g. by reducing the success rate of translation-invariant ensemble attack from 82.6% to 31.9% in comparison to previous state-of-the-art). It can be deployed as a plug-and-play solution to protect a variety of vision systems, as we demonstrate for the case of classification, segmentation and detection.

Referring expression comprehension aims to localize the object instance described by a natural language expression. Current referring expression methods have achieved good performance. However, none of them is able to achieve real-time inference without accuracy drop. The reason for the relatively slow inference speed is that these methods artificially split the referring expression comprehension into two sequential stages including proposal generation and proposal ranking. It does not exactly conform to the habit of human cognition. To this end, we propose a novel Realtime Cross-modality Correlation Filtering method (RCCF). RCCF reformulates the referring expression comprehension as a correlation filtering process. The expression is first mapped from the language domain to the visual domain and then treated as a template (kernel) to perform correlation filtering on the image feature map. The peak value in the correlation heatmap indicates the center points of the target box. In addition, RCCF also regresses a 2-D object size and 2-D offset. The center point coordinates, object size and center point offset together to form the target bounding box. Our method runs at 40 FPS while achieving leading performance in RefClef, RefCOCO, RefCOCO+ and RefCOCOg benchmarks. In the challenging RefClef dataset, our methods almost double the state-of-the-art performance (34.70% increased to 63.79%). We hope this work can arouse more attention and studies to the new cross-modality correlation filtering framework as well as the one-stage framework for referring expression comprehension.

Modern adiabatic quantum computers (AQC) are already used to solve difficult combinatorial optimisation problems in various domains of science. Currently, only a few applications of AQC in computer vision have been demonstrated. We review AQC and derive a new algorithm for correspondence problems on point sets suitable for execution on AQC. Our algorithm has a subquadratic computational complexity of the state preparation. Examples of successful transformation estimation and point set alignment by simulated sampling are shown in the systematic experimental evaluation. Finally, we analyse the differences in the solutions and the corresponding energy values.

Even as deep neural networks have become very effective for tasks in vision and perception, it remains difficult to explain and debug their behavior. In this paper, we present a programmatic and semantic approach to explaining, understanding, and debugging the correct and incorrect behaviors of a neural network based perception system. Our approach is semantic in that it employs a high-level representation of the distribution of environment scenarios that the detector is intended to work on. It is programmatic in that the representation is a program in a domain-specific probabilistic programming language using which synthetic data can be generated to train and test the neural network. We present a framework that assesses the performance of the neural network to identify correct and incorrect detections, extracts rules from those results that semantically characterizes the correct and incorrect scenarios, and then specializes the probabilistic program with those rules in order to more precisely characterize the scenarios in which the neural network operates correctly or not, without human intervention. We demonstrate our results using the Scenic probabilistic programming language and a neural network-based object detector. Our experiments show that it is possible to automatically generate compact rules that significantly increase the correct detection rate (or conversely the incorrect detection rate) of the network and can thus help with debugging and understanding its behavior.

Lacking rich and realistic data, learned single image denoising algorithms generalize poorly in real raw images that not resemble the data used for training. Although the problem can be alleviated by the heteroscedastic Gaussian noise model, the noise sources caused by digital camera electronics are still largely overlooked, despite their significant effect on raw measurement, especially under extremely low-light condition. To address this issue, we present a highly accurate noise formation model based on the characteristics of CMOS photosensors, thereby enabling us to synthesize realistic samples that better match the physics of image formation process. Given the proposed noise model, we additionally propose a method to calibrate the noise parameters for available modern digital cameras, which is simple and reproducible for any new device. We systematically study the generalizability of a neural network trained with existing schemes, by introducing a new low-light denoising dataset that covers many modern digital cameras from diverse brands. Extensive empirical results collectively show that by utilizing our proposed noise formation model, a network can reach the capability as if it had been trained with rich real data, which demonstrates the effectiveness of our noise formation model.

A great deal of research has demonstrated recently that multi-view stereo (MVS) matching can be solved with deep learning methods. However, these efforts were focused on close-range objects and only a very few of the deep learning-based methods were specifically designed for large-scale 3D urban reconstruction due to the lack of multi-view aerial image benchmarks. In this paper, we present a synthetic aerial dataset, called the WHU dataset, we created for MVS tasks, which, to our knowledge, is the first large-scale multi-view aerial dataset. It was generated from a highly accurate 3D digital surface model produced from thousands of real aerial images with precise camera parameters. We also introduce in this paper a novel network, called RED-Net, for wide-range depth inference, which we developed from a recurrent encoder-decoder structure to regularize cost maps across depths and a 2D fully convolutional network as framework. RED-Net's low memory requirements and high performance make it suitable for large-scale and highly accurate 3D Earth surface reconstruction. Our experiments confirmed that not only did our method exceed the current state-of-the-art MVS methods by more than 50% mean absolute error (MAE) with less memory and computational cost, but its efficiency as well. It outperformed one of the best commercial software programs based on conventional methods, improving their efficiency 16 times over. Moreover, we proved that our RED-Net model pre-trained on the synthetic WHU dataset can be efficiently transferred to very different multi-view aerial image datasets without any fine-tuning. Dataset and code are available at http://gpcv.whu.edu.cn/data.

We present a novel Relightable Neural Renderer (RNR) for simultaneous view synthesis and relighting using multi-view image inputs. Existing neural rendering (NR) does not explicitly model the physical rendering process and hence has limited capabilities on relighting. RNR instead models image formation in terms of environment lighting, object intrinsic attributes, and light transport function (LTF), each corresponding to a learnable component. In particular, the incorporation of a physically based rendering process not only enables relighting but also improves the quality of view synthesis. Comprehensive experiments on synthetic and real data show that RNR provides a practical and effective solution for conducting free-viewpoint relighting.

Training competitive deep video models is an order of magnitude slower than training their counterpart image models. Slow training causes long research cycles, which hinders progress in video understanding research. Following standard practice for training image models, video model training has used a fixed mini-batch shape: a specific number of clips, frames, and spatial size. However, what is the optimal shape? High resolution models perform well, but train slowly. Low resolution models train faster, but are less accurate. Inspired by multigrid methods in numerical optimization, we propose to use variable mini-batch shapes with different spatial-temporal resolutions that are varied according to a schedule. The different shapes arise from resampling the training data on multiple sampling grids. Training is accelerated by scaling up the mini-batch size and learning rate when shrinking the other dimensions. We empirically demonstrate a general and robust grid schedule that yields a significant out-of-the-box training speedup without a loss in accuracy for different models (I3D, non-local, SlowFast), datasets (Kinetics, Something-Something, Charades), and training settings (with and without pre-training, 128 GPUs or 1 GPU). As an illustrative example, the proposed multigrid method trains a ResNet-50 SlowFast network 4.5x faster (wall-clock time, same hardware) while also improving accuracy (+0.8% absolute) on Kinetics-400 compared to baseline training. Code is available online.

Existing shadow detection methods suffer from an intrinsic limitation in relying on limited labeled datasets, and they may produce poor results in some complicated situations. To boost the shadow detection performance, this paper presents a multi-task mean teacher model for semi-supervised shadow detection by leveraging unlabeled data and exploring the learning of multiple information of shadows simultaneously. To be specific, we first build a multi-task baseline model to simultaneously detect shadow regions, shadow edges, and shadow count by leveraging their complementary information and assign this baseline model to the student and teacher network. After that, we encourage the predictions of the three tasks from the student and teacher networks to be consistent for computing a consistency loss on unlabeled data, which is then added to the supervised loss on the labeled data from the predictions of the multi-task baseline model. Experimental results on three widely-used benchmark datasets show that our method consistently outperforms all the compared state-of- the-art methods, which verifies that the proposed network can effectively leverage additional unlabeled data to boost the shadow detection performance.

Contemporary approaches frame the color constancy problem as learning camera specific illuminant mappings. While high accuracy can be achieved on camera specific data, these models depend on camera spectral sensitivity and typically exhibit poor generalisation to new devices. Additionally, regression methods produce point estimates that do not explicitly account for potential ambiguities among plausible illuminant solutions, due to the ill-posed nature of the problem. We propose a Bayesian framework that naturally handles color constancy ambiguity via a multi-hypothesis strategy. Firstly, we select a set of candidate scene illuminants in a data-driven fashion and apply them to a target image to generate a set of corrected images. Secondly, we estimate, for each corrected image, the likelihood of the light source being achromatic using a camera-agnostic CNN. Finally, our method explicitly learns a final illumination estimate from the generated posterior probability distribution. Our likelihood estimator learns to answer a camera-agnostic question and thus enables effective multi-camera training by disentangling illuminant estimation from the supervised learning task. We extensively evaluate our proposed approach and additionally set a benchmark for novel sensor generalisation without re-training. Our method provides state-of-the-art accuracy on multiple public datasets (up to 11% median angular error improvement) while maintaining real-time execution.