CVPR 2022 论文列表

We propose a method for jointly estimating the 3D motion, 3D shape, and appearance of highly motion-blurred objects from a video. To this end, we model the blurred appearance of a fast moving object in a generative fashion by parametrizing its 3D position, rotation, velocity, acceleration, bounces, shape, and texture over the duration of a predefined time window spanning multiple frames. Using differentiable rendering, we are able to estimate all parameters by minimizing the pixel-wise reprojection error to the input video via backpropagating through a rendering pipeline that accounts for motion blur by averaging the graphics output over short time intervals. For that purpose, we also estimate the camera exposure gap time within the same optimization. To account for abrupt motion changes like bounces, we model the motion trajectory as a piece-wise polynomial, and we are able to estimate the specific time of the bounce at sub-frame accuracy. Experiments on established benchmark datasets demonstrate that our method outperforms previous methods for fast moving object deblurring and 3D reconstruction.

In light of the success of contrastive learning in the image domain, current self-supervised video representation learning methods usually employ contrastive loss to facilitate video representation learning. When naively pulling two augmented views of a video closer, the model however tends to learn the common static background as a shortcut but fails to capture the motion information, a phenomenon dubbed as background bias. Such bias makes the model suffer from weak generalization ability, leading to worse performance on downstream tasks such as action recognition. To alleviate such bias, we propose Foreground-background Merging (FAME) to deliberately compose the moving foreground region of the selected video onto the static background of others. Specifically, without any off-the-shelf detector, we extract the moving foreground out of background regions via the frame difference and color statistics, and shuffle the background regions among the videos. By leveraging the semantic consistency between the original clips and the fused ones, the model focuses more on the motion patterns and is debiased from the background shortcut. Extensive experiments demonstrate that FAME can effectively resist background cheating and thus achieve the state-of-the-art performance on downstream tasks across UCF101, HMDB51, and Diving48 datasets. The code and configurations are released at https://github.com/Mark12Ding/FAME.

Implicit neural representation (INR) has been successful in representing static images. Contemporary image-based INR, with the use of Fourier-based positional encoding, can be viewed as a mapping from sinusoidal patterns with different frequencies to image content. Inspired by that view, we hypothesize that it is possible to generate temporally varying content with a single image-based INR model by displacing its input sinusoidal patterns over time. By exploiting the relation between the phase information in sinusoidal functions and their displacements, we incorporate into the conventional image-based INR model a phase-varying positional encoding module, and couple it with a phase-shift generation module that determines the phase-shift values at each frame. The model is trained end-to-end on a video to jointly determine the phase-shift values at each time with the mapping from the phase-shifted sinusoidal functions to the corresponding frame, enabling an implicit video representation. Experiments on a wide range of videos suggest that such a model is capable of learning to interpret phase-varying positional embeddings into the corresponding time-varying content. More importantly, we found that the learned phase-shift vectors tend to capture meaningful temporal and motion information from the video. In particular, manipulating the phase-shift vectors induces meaningful changes in the temporal dynamics of the resulting video, enabling non-trivial temporal and motion editing effects such as temporal interpolation, motion magnification, motion smoothing, and video loop detection.

In this paper we present VDTTS, a Visually-Driven Text-to-Speech model. Motivated by dubbing, VDTTS takes advantage of video frames as an additional input alongside text, and generates speech that matches the video signal. We demonstrate how this allows VDTTS to, unlike plain TTS models, generate speech that not only has prosodic variations like natural pauses and pitch, but is also synchronized to the input video. Experimentally, we show our model produces well-synchronized outputs, approaching the video-speech synchronization quality of the ground-truth, on several challenging benchmarks including "in-the-wild" content from VoxCeleb2. Supplementary demo videos demonstrating video-speech synchronization, robustness to speaker ID swapping, and prosody, presented at the project page.

MonoScene proposes a 3D Semantic Scene Completion (SSC) framework, where the dense geometry and semantics of a scene are inferred from a single monocular RGB image. Different from the SSC literature, relying on 2.5 or 3D input, we solve the complex problem of 2D to 3D scene reconstruction while jointly inferring its semantics. Our framework relies on successive 2D and 3D UNets bridged by a novel 2D-3D features projection inspired by optics and introduces a 3D context relation prior to enforce spatio-semantic consistency. Along with architectural contributions, we introduce novel global scene and local frustums losses. Experiments show we outperform the literature on all metrics and datasets while hallucinating plausible scenery even beyond the camera field of view. Our code and trained models are available at https://github.com/cv-rits/MonoScene.

Due to the inherent ill-posed nature of 2D-3D projection, monocular 3D object detection lacks accurate depth recovery ability. Although the deep neural network (DNN) enables monocular depth-sensing from high-level learned features, the pixel-level cues are usually omitted due to the deep convolution mechanism. To benefit from both the powerful feature representation in DNN and pixel-level geometric constraints, we reformulate the monocular object depth estimation as a progressive refinement problem and propose a joint semantic and geometric cost volume to model the depth error. Specifically, we first leverage neural networks to learn the object position, dimension, and dense normalized 3D object coordinates. Based on the object depth, the dense coordinates patch together with the corresponding object features is reprojected to the image space to build a cost volume in a joint semantic and geometric error manner. The final depth is obtained by feeding the cost volume to a refinement network, where the distribution of semantic and geometric error is regularized by direct depth supervision. Through effectively mitigating depth error by the refinement framework, we achieve state-of-the-art results on both the KITTI and Waymo datasets.

Monocular 3D object detection has attracted great attention for its advantages in simplicity and cost. Due to the ill-posed 2D to 3D mapping essence from the monocular imaging process, monocular 3D object detection suffers from inaccurate depth estimation and thus has poor 3D detection results. To alleviate this problem, we propose to introduce the ground plane as a prior in the monocular 3d object detection. The ground plane prior serves as an additional geometric condition to the ill-posed mapping and an extra source in depth estimation. In this way, we can get a more accurate depth estimation from the ground. Meanwhile, to take full advantage of the ground plane prior, we propose a depth-align training strategy and a precise two-stage depth inference method tailored for the ground plane prior. It is worth noting that the introduced ground plane prior requires no extra data sources like LiDAR, stereo images, and depth information. Extensive experiments on the KITTI benchmark show that our method could achieve state-of-the-art results compared with other methods while maintaining a very fast speed. Our code, models, and training logs are available at https://github.com/cfzd/MonoGround.

Monocular 3D object detection is an important yet challenging task in autonomous driving. Some existing methods leverage depth information from an off-the-shelf depth estimator to assist 3D detection, but suffer from the additional computational burden and achieve limited performance caused by inaccurate depth priors. To alleviate this, we propose MonoDTR, a novel end-to-end depth-aware transformer network for monocular 3D object detection. It mainly consists of two components: (1) the Depth-Aware Feature Enhancement (DFE) module that implicitly learns depth-aware features with auxiliary supervision without requiring extra computation, and (2) the Depth-Aware Transformer (DTR) module that globally integrates context- and depth-aware features. Moreover, different from conventional pixel-wise positional encodings, we introduce a novel depth positional encoding (DPE) to inject depth positional hints into transformers. Our proposed depth-aware modules can be easily plugged into existing image-only monocular 3D object detectors to improve the performance. Extensive experiments on the KITTI dataset demonstrate that our approach outperforms previous state-of-the-art monocular-based methods and achieves real-time detection. Code is available at https://github.com/kuanchihhuang/MonoDTR.

Benefiting from the pioneering design of generic object detectors, significant achievements have been made in the field of face detection. Typically, the architectures of the backbone, feature pyramid layer, and detection head module within the face detector all assimilate the excellent experience from general object detectors. However, several effective methods, including label assignment and scale-level data augmentation strategy, fail to maintain consistent superiority when applying on the face detector directly. Concretely, the former strategy involves a vast body of hyper-parameters and the latter one suffers from the challenge of scale distribution bias between different detection tasks, which both limit their generalization abilities. Furthermore, in order to provide accurate face bounding boxes for facial down-stream tasks, the face detector imperatively requires the elimination of false alarms. As a result, practical solutions on label assignment, scale-level data augmentation, and reducing false alarms are necessary for advancing face detectors. In this paper, we focus on resolving three aforementioned challenges that exiting methods are difficult to finish off and present a novel face detector, termed MogFace. In our Mogface, three key components, Adaptive Online Incremental Anchor Mining Strategy, Selective Scale Enhancement Strategy and Hierarchical Context-Aware Module, are separately proposed to boost the performance of face detectors. Finally, to the best of our knowledge, our MogFace is the best face detector on the Wider Face leader-board, achieving all champions across different testing scenarios. The code is available at https://github.com/damo-cv/MogFace.

Perceiving the similarity between images has been a long-standing and fundamental problem underlying various visual generation tasks. Predominant approaches measure the inter-image distance by computing pointwise absolute deviations, which tends to estimate the median of instance distributions and leads to blurs and artifacts in the generated images. This paper presents MoNCE, a versatile metric that introduces image contrast to learn a calibrated metric for the perception of multifaceted inter-image distances. Unlike vanilla contrast which indiscriminately pushes negative samples from the anchor regardless of their similarity, we propose to re-weight the pushing force of negative samples adaptively according to their similarity to the anchor, which facilitates the contrastive learning from informative negative samples. Since multiple patch-level contrastive objectives are involved in image distance measurement, we introduce optimal transport in MoNCE to modulate the pushing force of negative samples collaboratively across multiple contrastive objectives. Extensive experiments over multiple image translation tasks show that the proposed MoNCE outperforms various prevailing metrics substantially.

Recent works in video prediction have mainly focused on passive forecasting and low-level action-conditional prediction, which sidesteps the learning of interaction between agents and objects. We introduce the task of semantic action-conditional video prediction, which uses semantic action labels to describe those interactions and can be regarded as an inverse problem of action recognition. The challenge of this new task primarily lies in how to effectively inform the model of semantic action information. Inspired by the idea of Mixture of Experts, we embody each abstract label by a structured combination of various visual concept learners and propose a novel video prediction model, Modular Action Concept Network (MAC). Our method is evaluated on two newly designed synthetic datasets, CLEVR-Building-Blocks and Sapien-Kitchen, and one real-world dataset called Tower-Creation. Extensive experiments demonstrate that MAC can correctly condition on given instructions and generate corresponding future frames without need of bounding boxes. We further show that the trained model can make out-of-distribution generalization, be quickly adapted to new object categories and exploit its learnt features for object detection, showing the progression towards higher-level cognitive abilities. More visualizations can be found at http://www.pair.toronto.edu/mac/.

Noise modeling and reduction are fundamental tasks in low-level computer vision. They are particularly important for smartphone cameras relying on small sensors that exhibit visually noticeable noise. There has recently been renewed interest in using data-driven approaches to improve camera noise models via neural networks. These data-driven approaches target noise present in the raw-sensor image before it has been processed by the camera's image signal processor (ISP). Modeling noise in the RAW-rgb domain is useful for improving and testing the in-camera denoising algorithm; however, there are situations where the camera's ISP does not apply denoising or additional denoising is desired when the RAW-rgb domain image is no longer available. In such cases, the sensor noise propagates through the ISP to the final rendered image encoded in standard RGB (sRGB). The nonlinear steps on the ISP culminate in a significantly more complex noise distribution in the sRGB domain and existing raw-domain noise models are unable to capture the sRGB noise distribution. We propose a new sRGB-domain noise model based on normalizing flows that is capable of learning the complex noise distribution found in sRGB images under various ISO levels. Our normalizing flows-based approach outperforms other models by a large margin in noise modeling and synthesis tasks. We also show that image denoisers trained on noisy images synthesized with our noise model outperforms those trained with noise from baselines models.

Text-based video segmentation aims to segment the target object in a video based on a describing sentence. Incorporating motion information from optical flow maps with appearance and linguistic modalities is crucial yet has been largely ignored by previous work. In this paper, we design a method to fuse and align appearance, motion, and linguistic features to achieve accurate segmentation. Specifically, we propose a multi-modal video transformer, which can fuse and aggregate multi-modal and temporal features between frames. Furthermore, we design a language-guided feature fusion module to progressively fuse appearance and motion features in each feature level with guidance from linguistic features. Finally, a multi-modal alignment loss is proposed to alleviate the semantic gap between features from different modalities. Extensive experiments on A2D Sentences and J-HMDB Sentences verify the performance and the generalization ability of our method compared to the state-of-the-art methods.

Recent advances in implicit neural representations and differentiable rendering make it possible to simultaneously recover the geometry and materials of an object from multi-view RGB images captured under unknown static illumination. Despite the promising results achieved, indirect illumination is rarely modeled in previous methods, as it requires expensive recursive path tracing which makes the inverse rendering computationally intractable. In this paper, we propose a novel approach to efficiently recovering spatially-varying indirect illumination. The key insight is that indirect illumination can be conveniently derived from the neural radiance field learned from input images instead of being estimated jointly with direct illumination and materials. By properly modeling the indirect illumination and visibility of direct illumination, interreflection- and shadow-free albedo can be recovered. The experiments on both synthetic and real data demonstrate the superior performance of our approach compared to previous work and its capability to synthesize realistic renderings under novel viewpoints and illumination. Our code and data are available at https://zju3dv.github.io/invrender/.

We present a method that achieves state-of-the-art results on challenging (few-shot) layout-to-image generation tasks by accurately modeling textures, structures and relationships contained in a complex scene. After compressing RGB images into patch tokens, we propose the Transformer with Focal Attention (TwFA) for exploring dependencies of object-to-object, object-to-patch and patch-to-patch. Compared to existing CNN-based and Transformer-based generation models that entangled modeling on pixel-level & patch-level and object-level & patch-level respectively, the proposed focal attention predicts the current patch token by only focusing on its highly-related tokens that specified by the spatial layout, thereby achieving disambiguation during training. Furthermore, the proposed TwFA largely increases the data efficiency during training, therefore we propose the first few-shot complex scene generation strategy based on the well-trained TwFA. Comprehensive experiments show the superiority of our method, which significantly increases both quantitative metrics and qualitative visual realism with respect to state-of-the-art CNN-based and transformer-based methods. Code is available at https://github.com/JohnDreamer/TwFA.

Group re-identification (GReID) attempts to correctly associate groups with the same members under different cameras. The main challenge is how to resist the membership and layout variations. Existing works attempt to incorporate layout modeling on the basis of appearance features to achieve robust group representations. However, layout ambiguity is introduced because these methods only consider the 2D layout on the imaging plane. In this paper, we overcome the above limitations by 3D layout modeling. Specifically, we propose a novel 3D transformer (3DT) that reconstructs the relative 3D layout relationship among members, then applies sampling and quantification to preset a series of layout tokens along three dimensions, and selects the corresponding tokens as layout features for each member. Furthermore, we build a synthetic GReID dataset, City1M, including 1.84M images, 45K persons and 11.5K groups with 3D annotations to alleviate data shortages and poor annotations. To the best of our knowledge, 3DT is the first work to address GReID with 3D perspective, and the City1M is the currently largest dataset. Several experiments show the superiority of our 3DT and City1M. Our project has been released on https://github.com/LinlyAC/City1M-dataset.

Fusion of multiple sensor modalities such as camera, Lidar, and Radar, which are commonly found on autonomous vehicles, not only allows for accurate detection but also robustifies perception against adverse weather conditions and individual sensor failures. Due to inherent sensor characteristics, Radar performs well under extreme weather conditions (snow, rain, fog) that significantly degrade camera and Lidar. Recently, a few works have developed vehicle detection methods fusing Lidar and Radar signals, i.e., MVDNet. However, these models are typically developed under the assumption that the models always have access to two error-free sensor streams. If one of the sensors is unavailable or missing, the model may fail catastrophically. To mitigate this problem, we propose the Self-Training Multimodal Vehicle Detection Network (ST-MVDNet) which leverages a Teacher-Student mutual learning framework and a simulated sensor noise model used in strong data augmentation for Lidar and Radar. We show that by (1) enforcing output consistency between a Teacher network and a Student network and by (2) introducing missing modalities (strong augmentations) during training, our learned model breaks away from the error-free sensor assumption. This consistency enforcement enables the Student model to handle missing data properly and improve the Teacher model by updating it with the Student model's exponential moving average. Our experiments demonstrate that our proposed learning framework for multi-modal detection is able to better handle missing sensor data during inference. Furthermore, our method achieves new state-of-the-art performance (5% gain) on the Oxford Radar Robotcar dataset under various evaluation settings.

We present Mobile-Former, a parallel design of MobileNet and transformer with a two-way bridge in between. This structure leverages the advantages of MobileNet at local processing and transformer at global interaction. And the bridge enables bidirectional fusion of local and global features. Different from recent works on vision transformer, the transformer in Mobile-Former contains very few tokens (e.g. 6 or fewer tokens) that are randomly initialized to learn global priors, resulting in low computational cost. Combining with the proposed light-weight cross attention to model the bridge, Mobile-Former is not only computationally efficient, but also has more representation power. It outperforms MobileNetV3 at low FLOP regime from 25M to 500M FLOPs on ImageNet classification. For instance, Mobile-Former achieves 77.9% top-1 accuracy at 294M FLOPs, gaining 1.3% over MobileNetV3 but saving 17% of computations. When transferring to object detection, Mobile-Former outperforms MobileNetV3 by 8.6 AP in RetinaNet framework. Furthermore, we build an efficient end-to-end detector by replacing backbone, encoder and decoder in DETR with Mobile-Former, which outperforms DETR by 1.3 AP but saves 52% of computational cost and 36% of parameters. Code will be released at https://github.com/aaboys/mobileformer.

In this work, we propose a framework for single-view hand mesh reconstruction, which can simultaneously achieve high reconstruction accuracy, fast inference speed, and temporal coherence. Specifically, for 2D encoding, we propose lightweight yet effective stacked structures. Regarding 3D decoding, we provide an efficient graph operator, namely depth-separable spiral convolution. Moreover, we present a novel feature lifting module for bridging the gap between 2D and 3D representations. This module begins with a map-based position regression (MapReg) block to integrate the merits of both heatmap encoding and position regression paradigms for improved 2D accuracy and temporal coherence. Furthermore, MapReg is followed by pose pooling and pose-to-vertex lifting approaches, which transform 2D pose encodings to semantic features of 3D vertices. Overall, our hand reconstruction framework, called MobRecon, comprises affordable computational costs and miniature model size, which reaches a high inference speed of 83FPS on Apple A14 CPU. Extensive experiments on popular datasets such as FreiHAND, RHD, and HO3Dv2 demonstrate that our MobRecon achieves superior performance on reconstruction accuracy and temporal coherence. Our code is publicly available at https://github.com/SeanChenxy/HandMesh.