CVPR 2021 论文列表

We investigate the problem of generating 3D meshes from single free-hand sketches, aiming at fast 3D modeling for novice users. It can be regarded as a single-view reconstruction problem, but with unique challenges, brought by the variation and conciseness of sketches. Ambiguities in poorly-drawn sketches could make it hard to determine how the sketched object is posed. In this paper, we address the importance of viewpoint specification for overcoming such ambiguities, and propose a novel view-aware generation approach. By explicitly conditioning the generation process on a given viewpoint, our method can generate plausible shapes automatically with predicted viewpoints, or with specified viewpoints to help users better express their intentions. Extensive evaluations on various datasets demonstrate the effectiveness of our view-aware design in solving sketch ambiguities and improving reconstruction quality.

The dense video captioning task aims to detect and describe a sequence of events in a video for detailed and coherent storytelling. Previous works mainly adopt a "detect-then-describe" framework, which firstly detects event proposals in the video and then generates descriptions for the detected events. However, the definitions of events are diverse which could be as simple as a single action or as complex as a set of events, depending on different semantic contexts. Therefore, directly detecting events based on video information is ill-defined and hurts the coherency and accuracy of generated dense captions. In this work, we reverse the predominant "detect-then-describe" fashion, proposing a top-down way to first generate paragraphs from a global view and then ground each event description to a video segment for detailed refinement. It is formulated as a Sketch, Ground, and Refine process (SGR). The sketch stage first generates a coarse-grained multi-sentence paragraph to describe the whole video, where each sentence is treated as an event and gets localised in the grounding stage. In the refining stage, we improve captioning quality via refinement-enhanced training and dual-path cross attention on both coarse-grained event captions and aligned event segments. The updated event caption can further adjust its segment boundaries. Our SGR model outperforms state-of-the-art methods on ActivityNet Captioning benchmark under traditional and story-oriented dense caption evaluations. Code will be released at github.com/bearcatt/SGR.

Detecting aligned 3D keypoints is essential under many scenarios such as object tracking, shape retrieval and robotics. However, it is generally hard to prepare a high-quality dataset for all types of objects due to the ambiguity of keypoint itself. Meanwhile, current unsupervised detectors are unable to generate aligned keypoints with good coverage. In this paper, we propose an unsupervised aligned keypoint detector, Skeleton Merger, which utilizes skeletons to reconstruct objects. It is based on an Autoencoder architecture. The encoder proposes keypoints and predicts activation strengths of edges between keypoints. The decoder performs uniform sampling on the skeleton and refines it into small point clouds with pointwise offsets. Then the activation strengths are applied and the sub-clouds are merged. Composite Chamfer Distance (CCD) is proposed as a distance between the input point cloud and the reconstruction composed of sub-clouds masked by activation strengths. We demonstrate that Skeleton Merger is capable of detecting semantically-rich salient keypoints with good alignment, and shows comparable performance to supervised methods on the KeypointNet dataset. It is also shown that the detector is robust to noise and subsampling. Our code is available at https://github.com/eliphatfs/SkeletonMerger.

We introduce RoboPose, a method to estimate the joint angles and the 6D camera-to-robot pose of a known articulated robot from a single RGB image. This is an important problem to grant mobile and itinerant autonomous systems the ability to interact with other robots using only visual information in non-instrumented environments, especially in the context of collaborative robotics. It is also challenging because robots have many degrees of freedom and an infinite space of possible configurations that often result in self-occlusions and depth ambiguities when imaged by a single camera. The contributions of this work are three-fold. First, we introduce a new render & compare approach for estimating the 6D pose and joint angles of an articulated robot that can be trained from synthetic data, generalizes to new unseen robot configurations at test time, and can be applied to a variety of robots. Second, we experimentally demonstrate the importance of the robot parametrization for the iterative pose updates and design a parametrization strategy that is independent of the robot structure. Finally, we show experimental results on existing benchmark datasets for four different robots and demonstrate that our method significantly outperforms the state of the art. Code and pre-trained models are available on the project webpage.

Existing methods for single-view 3D object reconstruction directly learn to transform image features into 3D representations. However, these methods are vulnerable to images containing noisy backgrounds and heavy occlusions because the extracted image features do not contain enough information to reconstruct high-quality 3D shapes. Humans routinely use incomplete or noisy visual cues from an image to retrieve similar 3D shapes from their memory and reconstruct the 3D shape of an object. Inspired by this, we propose a novel method, named Mem3D, that explicitly constructs shape priors to supplement the missing information in the image. Specifically, the shape priors are in the forms of "image-voxel" pairs in the memory network, which is stored by a well-designed writing strategy during training. We also propose a voxel triplet loss function that helps to retrieve the precise 3D shapes that are highly related to the input image from shape priors. The LSTM-based shape encoder is introduced to extract information from the retrieved 3D shapes, which are useful in recovering the 3D shape of an object that is heavily occluded or in complex environments. Experimental results demonstrate that Mem3D significantly improves reconstruction quality and performs favorably against state-of-the-art methods on the ShapeNet and Pix3D datasets.

Instance shadow detection aims to find shadow instances paired with the objects that cast the shadows. The previous work adopts a two-stage framework to first predict shadow instances, object instances, and shadow-object associations from the region proposals, then leverage a post-processing to match the predictions to form the final shadow-object pairs. In this paper, we present a new single-stage fully-convolutional network architecture with a bidirectional relation learning module to directly learn the relations of shadow and object instances in an end-to-end manner. Compared with the prior work, our method actively explores the internal relationship between shadows and objects to learn a better pairing between them, thus improving the overall performance for instance shadow detection. We evaluate our method on the benchmark dataset for instance shadow detection, both quantitatively and visually. The experimental results demonstrate that our method clearly outperforms the state-of-the-art method.

The task of motion transfer between a source dancer and a target person is a special case of the pose transfer problem, in which the target person changes their pose in accordance with the motions of the dancer. In this work, we propose a novel method that can reanimate a single image by arbitrary video sequences, unseen during training. The method combines three networks: (i) a segmentation-mapping network, (ii) a realistic frame-rendering network, and (iii) a face refinement network. By separating this task into three stages, we are able to attain a novel sequence of realistic frames, capturing natural motion and appearance. Our method obtains significantly better visual quality than previous methods and is able to animate diverse body types and appearances, which are captured in challenging poses.

Non-visual imaging sensors are widely used in the industry for different purposes. Those sensors are more expensive than visual (RGB) sensors, and usually produce images with lower resolution. To this end, Cross-Modality Super-Resolution methods were introduced, where an RGB image of a high-resolution assists in increasing the resolution of a low-resolution modality. However, fusing images from different modalities is not a trivial task, since each multi-modal pair varies greatly in its internal correlations. For this reason, traditional state-of-the-arts which are trained on external datasets often struggle with yielding an artifact-free result that is still loyal to the target modality characteristics. We present CMSR, a single-pair approach for Cross-Modality Super-Resolution. The network is internally trained on the two input images only, in a self-supervised manner, learns their internal statistics and correlations, and applies them to upsample the target modality. CMSR contains an internal transformer which is trained on-the-fly together with the up-sampling process itself and without supervision, to allow dealing with pairs that are only weakly aligned. We show that CMSR produces state-of-the-art super resolved images, yet without introducing artifacts or irrelevant details that originate from the RGB image only.

In this paper, we consider the absorption effect for the problem of single image reflection removal. We show that the absorption effect can be numerically approximated by the average of refractive amplitude coefficient map. We then reformulate the image formation model and propose a two-step solution that explicitly takes the absorption effect into account. The first step estimates the absorption effect from a reflection-contaminated image, while the second step recovers the transmission image by taking a reflection-contaminated image and the estimated absorption effect as the input. Experimental results on four public datasets show that our two-step solution not only successfully removes reflection artifact, but also faithfully restores the intensity distortion caused by the absorption effect. Our ablation studies further demonstrate that our method achieves superior performance on the recovery of overall intensity and has good model generalization capacity. The code is available at https://github.com/q-zh/absorption.

We present a novel method for predicting accurate depths from monocular images with high efficiency. This optimal efficiency is achieved by exploiting wavelet decomposition, which is integrated in a fully differentiable encoder-decoder architecture. We demonstrate that we can reconstruct high-fidelity depth maps by predicting sparse wavelet coefficients. In contrast with previous works, we show that wavelet coefficients can be learned without direct supervision on coefficients. Instead we supervise only the final depth image that is reconstructed through the inverse wavelet transform. We additionally show that wavelet coefficients can be learned in fully self-supervised scenarios, without access to ground-truth depth. Finally, we apply our method to different state-of-the-art monocular depth estimation models, in each case giving similar or better results compared to the original model, while requiring less than half the multiply-adds in the decoder network.

Camouflage is a key defence mechanism across species that is critical to survival. Common camouflage include background matching, imitating the color and pattern of the environment, and disruptive coloration, disguising body outlines. Camouflaged object detection (COD) aims to segment camouflaged objects hiding in their surroundings. Existing COD models are built upon binary ground truth to segment the camouflaged objects without illustrating the level of camouflage. In this paper, we revisit this task and argue that explicitly modeling the conspicuousness of camouflaged objects against their particular backgrounds can not only lead to a better understanding about camouflage and evolution of animals, but also provide guidance to design more sophisticated camouflage techniques. Furthermore, we observe that it is some specific parts of the camouflaged objects that make them detectable by predators. With the above understanding about camouflaged objects, we present the first ranking based COD network to simultaneously localize, segment and rank camouflaged objects. The localization model is proposed to find the discriminative regions that make the camouflaged object obvious. The segmentation model segments the full scope of the camouflaged objects. And, the ranking model infers the detectability of different camouflaged objects. Moreover, we contribute a large COD testing set to evaluate the generalization ability of COD models. Experimental results show that our model achieves new state-of-the-art, leading to a more interpretable COD network.

Many adversarial attacks have been proposed to investigate the security issues of deep neural networks. In the black-box setting, current model stealing attacks train a substitute model to counterfeit the functionality of the target model. However, the training requires querying the target model. Consequently, the query complexity remains high, and such attacks can be defended easily. This study aims to train a generalized substitute model called "Simulator", which can mimic the functionality of any unknown target model. To this end, we build the training data with the form of multiple tasks by collecting query sequences generated during the attacks of various existing networks. The learning process uses a mean square error-based knowledge-distillation loss in the meta-learning to minimize the difference between the Simulator and the sampled networks. The meta-gradients of this loss are then computed and accumulated from multiple tasks to update the Simulator and subsequently improve generalization. When attacking a target model that is unseen in training, the trained Simulator can accurately simulate its functionality using its limited feedback. As a result, a large fraction of queries can be transferred to the Simulator, thereby reducing query complexity. Results of the comprehensive experiments conducted using the CIFAR-10, CIFAR-100, and TinyImageNet datasets demonstrate that the proposed approach reduces query complexity by several orders of magnitude compared to the baseline method. The implementation source code is released online at https://github.com/machanic/SimulatorAttack.

One strategy for adversarially training a robust model is to maximize its certified radius -- the neighborhood around a given training sample for which the model's prediction remains unchanged. The scheme typically involves analyzing a "smoothed" classifier where one estimates the prediction corresponding to Gaussian samples in the neighborhood of each sample in the mini-batch, accomplished in practice by Monte Carlo sampling. In this paper, we investigate the hypothesis that this sampling bottleneck can potentially be mitigated by identifying ways to directly propagate the covariance matrix of the smoothed distribution through the network. To this end, we find that other than certain adjustments to the network, propagating the covariances must also be accompanied by additional accounting that keeps track of how the distributional moments transform and interact at each stage in the network. We show how satisfying these criteria yields an algorithm for maximizing the certified radius on datasets including Cifar-10, ImageNet, and Places365 while offering runtime savings on networks with moderate depth, with a small compromise in overall accuracy. We describe the details of the key modifications that enable practical use. Via various experiments, we evaluate when our simplifications are sensible, and what the key benefits and limitations are.

Building instance segmentation models that are data-efficient and can handle rare object categories is an important challenge in computer vision. Leveraging data augmentations is a promising direction towards addressing this challenge. Here, we perform a systematic study of the Copy-Paste augmentation (e.g., [13, 12]) for instance segmentation where we randomly paste objects onto an image. Prior studies on Copy-Paste relied on modeling the surrounding visual context for pasting the objects. However, we find that the simple mechanism of pasting objects randomly is good enough and can provide solid gains on top of strong baselines. Furthermore, we show Copy-Paste is additive with semi-supervised methods that leverage extra data through pseudo labeling (eg. self-training). On COCO instance segmentation, we achieve 49.1 mask AP and 57.3 box AP, an improvement of +0.6 mask AP and +1.5 box AP over the previous state-of-the-art. We further demonstrate that Copy-Paste can lead to significant improvements on the LVIS benchmark. Our baseline model outperforms the LVIS 2020 Challenge winning entry by +3.6 mask AP on rare categories.

Accurate estimation of 3D human motion from monocular video requires modeling both kinematics (body motion without physical forces) and dynamics (motion with physical forces). To demonstrate this, we present SimPoE, a Simulation-based approach for 3D human Pose Estimation, which integrates image-based kinematic inference and physics-based dynamics modeling. SimPoE learns a policy that takes as input the current-frame pose estimate and the next image frame to control a physically-simulated character to output the next-frame pose estimate. The policy contains a learnable kinematic pose refinement unit that uses 2D keypoints to iteratively refine its kinematic pose estimate of the next frame. Based on this refined kinematic pose, the policy learns to compute dynamics-based control (e.g., joint torques) of the character to advance the current-frame pose estimate to the pose estimate of the next frame. This design couples the kinematic pose refinement unit with the dynamics-based control generation unit, which are learned jointly with reinforcement learning to achieve accurate and physically-plausible pose estimation. Furthermore, we propose a meta-control mechanism that dynamically adjusts the character's dynamics parameters based on the character state to attain more accurate pose estimates. Experiments on large-scale motion datasets demonstrate that our approach establishes the new state of the art in pose accuracy while ensuring physical plausibility.

A common classification task situation is where one has a large amount of data available for training, but only a small portion is annotated with class labels. The goal of semi-supervised training, in this context, is to improve classification accuracy by leverage information not only from labeled data but also from a large amount of unlabeled data. Recent works have developed significant improvements by exploring the consistency constrain between differently augmented labeled and unlabeled data. Following this path, we propose a novel unsupervised objective that focuses on the less studied relationship between the high confidence unlabeled data that are similar to each other. The new proposed Pair Loss minimizes the statistical distance between high confidence pseudo labels with similarity above a certain threshold. Combining the Pair Loss with the techniques developed by the MixMatch family, our proposed SimPLE algorithm shows significant performance gains over previous algorithms on CIFAR-100 and Mini-ImageNet, and is on par with the state-of-the-art methods on CIFAR-10 and SVHN. Furthermore, SimPLE also outperforms the state-of-the-art methods in the transfer learning setting, where models are initialized by the weights pre-trained on ImageNet or DomainNet-Real. The code is available at github.com/zijian-hu/SimPLE.

Shape modeling and reconstruction from raw point clouds of objects stand as a fundamental challenge in vision and graphics research. Classical methods consider analytic shape priors; however, their performance is degraded when the scanned points deviate from the ideal conditions of cleanness and completeness. Important progress has been recently made by data-driven approaches, which learn global and/or local models of implicit surface representations from auxiliary sets of training shapes. Motivated from a universal phenomenon that self-similar shape patterns of local surface patches repeat across the entire surface of an object, we aim to push forward the data-driven strategies and propose to learn a local implicit surface network for a shared, adaptive modeling of the entire surface for a direct surface reconstruction from raw point cloud; we also enhance the leveraging of surface self-similarities by improving correlations among the optimized latent codes of individual surface patches. Given that orientations of raw points could be unavailable or noisy, we extend signagnostic learning into our local implicit model, which enables our recovery of signed implicit fields of local surfaces from the unsigned inputs. We term our framework as Sign-Agnostic Implicit Learning of Surface Self-Similarities (SAIL-S3). With a global post-optimization of local sign flipping, SAIL-S3 is able to directly model raw, un-oriented point clouds and reconstruct high-quality object surfaces. Experiments show its superiority over existing methods.

We propose a novel Siamese Natural Language Tracker (SNLT), which brings the advancements in visual tracking to the tracking by natural language (NL) specification task. The proposed SNLT is applicable to a wide range of Siamese trackers, providing a new class of baselines for the tracking by NL task and promising future improvements from the advancements of Siamese trackers. The carefully designed architecture of the Siamese Natural Language Region Proposal Network (SNL-RPN), together with the Dynamic Aggregation of vision and language modalities, is introduced to perform the tracking by NL task. Empirical results over tracking benchmarks with NL annotations show that the proposed SNLT improves Siamese trackers by 3 to 7 percentage points with a slight tradeoff of speed. The proposed SNLT outperforms all NL trackers to-date and is competitive among state-of-the-art real-time trackers on LaSOT benchmarks while running at 50 frames per second on a single GPU.

In this work, we focus on improving online multi-object tracking (MOT). In particular, we propose a novel region-based Siamese Multi-Object Tracking network, which we name SiamMOT. SiamMOT is based upon Faster-RCNN and adds a forward tracker that models the instance's motion across two frames such that detected instances can be associated in an online fashion. We present two variants of this tracker, an implicit motion model and a novel Siamese-type explicit motion model. We carry out extensive quantitative experiments on three important MOT datasets: MOT17, TAO-person and Caltech Roadside Pedestrians, showing the importance of motion modelling for MOT and the ability of SiamMOT to substantially outperform the state-of-the-art. Finally, SiamMOT also outperforms the winners of ACM MM'20 HiEve Grand Challenge on the Human in Events dataset. Moreover, SiamMOT is efficient, and it runs at 17 FPS for 720P videos on a single modern GPU. We will release SiamMOT source code upon acceptance of this paper.