ICCV 2023 Papers

We present CaPhy, a novel method for reconstructing animatable human avatars with realistic dynamic properties for clothing. Specifically, we aim for capturing the geometric and physical properties of the clothing from real observations. This allows us to apply novel poses to the human avatar with physically correct deformations and wrinkles of the clothing. To this end, we combine unsupervised training with physics-based losses and 3D-supervised training using scanned data to reconstruct a dynamic model of clothing that is physically realistic and conforms to the human scans. We also optimize the physical parameters of the underlying physical model from the scans by introducing gradient constraints of the physics-based losses. In contrast to previous work on 3D avatar reconstruction, our method is able to generalize to novel poses with realistic dynamic cloth deformations. Experiments on several subjects demonstrate that our method can estimate the physical properties of the garments, resulting in superior quantitative and qualitative results compared with previous methods.

Semantic scene completion (SSC) requires an accurate understanding of the geometric and semantic relationships between the objects in the 3D scene for reasoning the occluded objects. The popular SSC methods voxelize the 3D objects, allowing the deep 3D convolutional network (3D CNN) to learn the object relationships from the complex scenes. However, the current networks lack the controllable kernels to model the object relationship across multiple views, where appropriate views provide the relevant information for suggesting the existence of the occluded objects. In this paper, we propose Cross-View Synthesis Transformer (CVSformer), which consists of Multi-View Feature Synthesis and Cross-View Transformer for learning cross-view object relationships. In the multi-view feature synthesis, we use a set of 3D convolutional kernels rotated differently to compute the multi-view features for each voxel. In the cross-view transformer, we employ the cross-view fusion to comprehensively learn the cross-view relationships, which form useful information for enhancing the features of individual views. We use the enhanced features to predict the geometric occupancies and semantic labels of all voxels. We evaluate CVSformer on public datasets, where CVSformer yields state-of-the-art results. Our code is available at https://github.com/donghaotian123/CVSformer.

Recent advances in neural reconstruction using posed image sequences have made remarkable progress. However, due to the lack of depth information, existing volumetric-based techniques simply duplicate 2D image features of the object surface along the entire camera ray. We contend this duplication introduces noise in empty and occluded spaces, posing challenges for producing high-quality 3D geometry. Drawing inspiration from traditional multi-view stereo methods, we propose an end-to-end 3D neural reconstruction framework CVRecon, designed to exploit the rich geometric embedding in the cost volumes to facilitate 3D geometric feature learning. Furthermore, we present Ray-contextual Compensated Cost Volume (RCCV), a novel 3D geometric feature representation that encodes view-dependent information with improved integrity and robustness. Through comprehensive experiments, we demonstrate that our approach significantly improves the reconstruction quality in various metrics and recovers clear fine details of the 3D geometries. Our extensive ablation studies provide insights into the development of effective 3D geometric feature learning schemes. Project page: https://cvrecon.ziyue.cool

The discrimination of instance embeddings plays a vital role in associating instances across time for online video instance segmentation (VIS). Instance embedding learning is directly supervised by the contrastive loss computed upon the contrastive items (CIs), which are sets of anchor/positive/negative embeddings. Recent online VIS methods leverage CIs sourced from one reference frame only, which we argue is insufficient for learning highly discriminative embeddings. Intuitively, a possible strategy to enhance CIs is replicating the inference phase during training. To this end, we propose a simple yet effective training strategy, called Consistent Training for Online VIS(CTVIS), which devotes to aligning the training and inference pipelines in terms of building CIs. Specifically, CTVIS constructs CIs by referring inference the momentum-averaged embedding and the memory bank storage mechanisms, and adding noise to the relevant embeddings. Such an extension allows a reliable comparison between embeddings of current instances and the stable representations of historical instances, thereby conferring an advantage in modeling VIS challenges such as occlusion, re-identification, and deformation. Empirically, CTVIS outstrips the SOTA VIS models by up to +5.0 points on three VIS benchmarks, including YTVIS19 (55.1% AP), YTVIS21 (50.1% AP) and OVIS (35.5% AP). Furthermore, we find that pseudo-videos transformed from images can train robust models surpassing fully-supervised ones.

Vision-Language Pretraining (VLP) has shown impressive results on diverse downstream tasks by offline training on large-scale datasets. Regarding the growing nature of real-world data, such an offline training paradigm on ever-expanding data is unsustainable, because models lack the continual learning ability to accumulate knowledge constantly. However, most continual learning studies are limited to uni-modal classification and existing multi-modal datasets cannot simulate continual non-stationary data stream scenarios. To support the study of Vision-Language Continual Pretraining (VLCP), we first contribute a comprehensive and unified benchmark dataset P9D which contains over one million product image-text pairs from 9 industries. The data from each industry as an independent task supports continual learning and conforms to the real-world long-tail nature to simulate pretraining on web data. We comprehensively study the characteristics and challenges of VLCP, and propose a new algorithm: Compatible momentum contrast with Topology Preservation, dubbed CTP. The compatible momentum model absorbs the knowledge of the current and previous-task models to flexibly update the modal feature. Moreover, Topology Preservation transfers the knowledge of embedding across tasks while preserving the flexibility of feature adjustment. The experimental results demonstrate our method not only achieves superior performance compared with other baselines but also does not bring an expensive training burden. Dataset and codes are available at https://github.com/KevinLight831/CTP.

Domain Adaptive Object Detection (DAOD) aims to improve the detection performance of target domains by minimizing the feature distribution between the source and target domain. Recent approaches usually align such distributions in terms of categories through adversarial learning and some progress has been made. However, when objects are non-uniformly distributed at different scales, such category-level alignment causes imbalanced object feature learning, refer as the inconsistency of category alignment at different scales. For better category-level feature alignment, we propose a novel DAOD framework of joint category and scale information, dubbed CSDA, such a design enables effective object learning for different scales. Specifically, our framework is implemented by two closely-related modules: 1) SGFF (Scale-Guided Feature Fusion) fuses the category representations of different domains to learn category-specific features, where the features are aligned by discriminators at three scales. 2) SAFE (Scale-Auxiliary Feature Enhancement) encodes scale coordinates into a group of tokens and enhances the representation of category-specific features at different scales by self-attention. Based on the anchor-based Faster-RCNN and anchor-free FCOS detectors, experiments show that our method achieves state-of-the-art results on three DAOD benchmarks.

Beyond novel view synthesis, Neural Radiance Fields are useful for applications that interact with the real world. In this paper, we use them as an implicit map of a given scene and propose a camera relocalization algorithm tailored for this representation. The proposed method enables to compute in real-time the precise position of a device using a single RGB camera, during its navigation. In contrast with previous work, we do not rely on pose regression or photometric alignment but rather use dense local features obtained through volumetric rendering which are specialized on the scene with a self-supervised objective. As a result, our algorithm is more accurate than competitors, able to operate in dynamic outdoor environments with changing lightning conditions and can be readily integrated in any volumetric neural renderer.

Autonomous driving requires an accurate and fast 3D perception system that includes 3D object detection, tracking, and segmentation. Although recent low-cost camera-based approaches have shown promising results, they are susceptible to poor illumination or bad weather conditions and have a large localization error. Hence, fusing camera with low-cost radar, which provides precise long-range measurement and operates reliably in all environments, is promising but has not yet been thoroughly investigated. In this paper, we propose Camera Radar Net (CRN), a novel camera-radar fusion framework that generates a semantically rich and spatially accurate bird's-eye-view (BEV) feature map for various tasks. To overcome the lack of spatial information in an image, we transform perspective view image features to BEV with the help of sparse but accurate radar points. We further aggregate image and radar feature maps in BEV using multi-modal deformable attention designed to tackle the spatial misalignment between inputs. CRN with real-time setting operates at 20 FPS while achieving comparable performance to LiDAR detectors on nuScenes, and even outperforms at a far distance on 100m setting. Moreover, CRN with offline setting yields 62.4% NDS, 57.5% mAP on nuScenes test set and ranks first among all camera and camera-radar 3D object detectors.

We study the task of weakly-supervised point cloud semantic segmentation with sparse annotations (e.g., less than 0.1% points are labeled), aiming to reduce the expensive cost of dense annotations. Unfortunately, with extremely sparse annotated points, it is very difficult to extract both contextual and object information for scene understanding such as semantic segmentation. Motivated by masked modeling (e.g., MAE) in image and video representation learning, we seek to endow the power of masked modeling to learn contextual information from sparsely-annotated points. However, directly applying MAE to 3D point clouds with sparse annotations may fail to work. First, it is nontrivial to effectively mask out the informative visual context from 3D point clouds. Second, how to fully exploit the sparse annotations for context modeling remains an open question. In this paper, we propose a simple yet effective Contextual Point Cloud Modeling (CPCM) method that consists of two parts: a region-wise masking (RegionMask) strategy and a contextual masked training (CMT) method. Specifically, RegionMask masks the point cloud continuously in geometric space to construct a meaningful masked prediction task for subsequent context learning. CMT disentangles the learning of supervised segmentation and unsupervised masked context prediction for effectively learning the very limited labeled points and mass unlabeled points, respectively. Extensive experiments on the widely-tested ScanNet V2 and S3DIS benchmarks demonstrate the superiority of CPCM over the state-of-the-art.

This paper presents CORE, a conceptually simple, effective and communication-efficient model for multi-agent cooperative perception. It addresses the task from a novel perspective of cooperative reconstruction, based on two key insights: 1) cooperating agents together provide a more holistic observation of the environment, and 2) the holistic observation can serve as valuable supervision to explicitly guide the model learning how to reconstruct the ideal observation based on collaboration. CORE instantiates the idea with three major components: a compressor for each agent to create more compact feature representation for efficient broadcasting, a lightweight attentive collaboration component for cross-agent message aggregation, and a reconstruction module to reconstruct the observation based on aggregated feature representations. This learning-to-reconstruct idea is task-agnostic, and offers clear and reasonable supervision to inspire more effective collaboration, eventually promoting perception tasks. We validate CORE on two large-scale multi-agent percetion dataset, OPV2V and V2X-Sim, in two tasks, i.e., 3D object detection and semantic segmentation. Results demonstrate that CORE achieves state-of-the-art performance, and is more communication-efficient.

The ill-posed nature of monocular 3D geometry (depth map and surface normals) estimation makes it rely mostly on data-driven approaches such as Deep Neural Networks (DNN). However, data acquisition of surface normals, especially the reliable normals, is acknowledged difficult. Commonly, reconstruction of surface normals with high quality is heuristic and time-consuming. Such fact urges methodologies to minimize dependency on ground-truth normals when predicting 3D geometry. In this work, we devise CO-planarity REgularized (CORE) loss functions and Structure-Aware Normal Estimator (SANE). Without involving any knowledge of ground-truth normals, these two designs enable pixel-wise 3D geometry estimation weakly supervised by only ground-truth depth map. For CORE loss functions, the key idea is to exploit locally linear depth-normal orthogonality under spherical coordinates as pixel-level constraints, and utilize our designed Adaptive Polar Regularization (APR) to resolve underlying numerical degeneracies. Meanwhile, SANE easily establishes multi-task learning with CORE loss functions on both depth and surface normal estimation, leading to the whole performance leap. Extensive experiments present the effectiveness of our method on various DNN architectures and data benchmarks. The experimental results demonstrate that our depth estimation achieves the state-of-the-art performance across all metrics on indoor scenes and comparable performance on outdoor scenes. In addition, our surface normal estimation is overall superior.

The ability to forecast human-environment collisions from egocentric observations is vital to enable collision avoidance in applications such as VR, AR, and wearable assistive robotics. In this work, we introduce the challenging problem of predicting collisions in diverse environments from multi-view egocentric videos captured from body-mounted cameras. Solving this problem requires a generalizable perception system that can classify which human body joints will collide and estimate a collision region heatmap to localize collisions in the environment. To achieve this, we propose a transformer-based model called COPILOT to perform collision prediction and localization simultaneously, which accumulates information across multi-view inputs through a novel 4D space-time-viewpoint attention mechanism. To train our model and enable future research on this task, we develop a synthetic data generation framework that produces egocentric videos of virtual humans moving and colliding within diverse 3D environments. This framework is then used to establish a large-scale dataset consisting of 8.6M egocentric RGBD frames. Extensive experiments show that COPILOT generalizes to unseen synthetic as well as real-world scenes. We further demonstrate COPILOT outputs are useful for downstream collision avoidance through simple closed-loop control. Please visit our project webpage at https://sites.google.com/stanford.edu/copilot.

Generating life-like whole-body human grasping has garnered significant attention in the field of computer graphics. Existing works have demonstrated the effectiveness of keyframe-guided motion generation framework, witch focus on modeling the grasping motions of humans in temporal sequence when the target objects are placed in front of them. However, the generated grasping poses of the human body in the key-frames are limited, failing to capture the full range of grasping poses that humans are capable of. To address this issue, we propose a novel framework called COOP (DeCOupling and COupling of Whole-Body GrasPing Pose Generation) to synthesize life-like whole-body poses that cover the widest range of human grasping capabilities. In this framework, we first decouple the whole-body pose into body pose and hand pose and model them separately, which allows us to pre-train the body model with out-of-domain data easily. Then, we couple these two generated body parts through a unified optimization algorithm. Furthermore, we design a simple evaluation method to evaluate the generalization ability of models in generating grasping poses for objects placed at different positions. The experimental results demonstrate the efficacy and superiority of our method. And COOP holds great potential as a plug-and-play component for other domains in whole-body pose generation. Our models and code are available at https://github.com/zhengyanzhao1997/COOP.

We introduce COOL-CHIC, a Coordinate-based Low Complexity Hierarchical Image Codec. It is a learned alternative to autoencoders with 629 parameters and 680 multiplications per decoded pixel. COOL-CHIC offers compression performance close to modern conventional MPEG codecs such as HEVC and is competitive with popular autoencoder-based systems. This method is inspired by Coordinate-based Neural Representations, where an image is represented as a learned function which maps pixel coordinates to RGB values. The parameters of the mapping function are then sent using entropy coding. At the receiver side, the compressed image is obtained by evaluating the mapping function for all pixel coordinates. COOL-CHIC implementation is made open-source.

Recently, neural network (NN)-based image compression studies have actively been made and has shown impressive performance in comparison to traditional methods. However, most of the works have focused on non-scalable image compression (single-layer coding) while spatially scalable image compression has drawn less attention although it has many applications. In this paper, we propose a novel NN-based spatially scalable image compression method, called COMPASS, which supports arbitrary-scale spatial scalability. Our proposed COMPASS has a very flexible structure where the number of layers and their respective scale factors can be arbitrarily determined during inference. To reduce the spatial redundancy between adjacent layers for arbitrary scale factors, our COMPASS adopts an inter-layer arbitrary scale prediction method, called LIFF, based on implicit neural representation. We propose a combined RD loss function to effectively train multiple layers. Experimental results show that our COMPASS achieves BD-rate gain of -58.33% and -47.17% at maximum compared to SHVC and the state-of-the-art NN-based spatially scalable image compression method, respectively, for various combinations of scale factors. Our COMPASS also shows comparable or even better coding efficiency than the single-layer coding for various scale factors.

Practical object detection application can lose its effectiveness on image inputs with natural distribution shifts. This problem leads the research community to pay more attention on the robustness of detectors under Out-Of-Distribution (OOD) inputs. Existing works construct datasets to benchmark the detector's OOD robustness for a specific application scenario, e.g., Autonomous Driving. However, these datasets lack universality and are hard to benchmark general detectors built on common tasks such as COCO. To give a more comprehensive robustness assessment, we introduce COCO-O(ut-of-distribution), a test dataset based on COCO with 6 types of natural distribution shifts. COCO-O has a large distribution gap with training data and results in a significant 55.7% relative performance drop on a Faster R-CNN detector. We leverage COCO-O to conduct experiments on more than 100 modern object detectors to investigate if their improvements are credible or just over-fitting to the COCO test set. Unfortunately, most classic detectors in early years do not exhibit strong OOD generalization. We further study the robustness effect on recent breakthroughs of detector's architecture design, augmentation and pre-training techniques. Some empirical findings are revealed: 1) Compared with detection head or neck, backbone is the most important part for robustness; 2) An end-to-end detection transformer design brings no enhancement, and may even reduce robustness; 3) Large-scale foundation models have made a great leap on robust object detection. We hope our COCO-O could provide a rich testbed for robustness study of object detection. The dataset will be available at https://github.com/alibaba/easyrobust/tree/main/benchmarks/coco_o.

Predicting survival rates based on multi-gigapixel histopathology images is one of the most challenging tasks in digital pathology. Due to the computational complexities, Multiple Instance Learning (MIL) has become the conventional approach for this process as it breaks the image into smaller patches. However, this technique fails to account for the individual cells present in each patch, while they are the fundamental part of the tissue. In this work, we developed a novel dynamic and hierarchical point-cloud-based method (CO-PILOT) for the processing of cellular graphs extracted from routine histopathology images. By using bottom-up information propagation and top-down conditional attention, our model gains access to an adaptive focus across different levels of tissue hierarchy. Through comprehensive experiments, we demonstrate that our model can outperform all the state-of-the-art methods in survival prediction, including the hierarchical Vision Transformer (ViT), across two datasets and four metrics with only half of the parameters of the closest baseline. Importantly, our model is able to stratify the patients into different risk cohorts with statistically different outcomes across two large datasets, a task that was previously achievable only using genomic information. Furthermore, we publish a large dataset containing 873 cellular graphs from 188 patients, along with their survival information, making it one of the largest publicly available datasets in this context.

We present CO-Net, a cohesive framework that optimizes multiple point cloud tasks collectively across heterogeneous dataset domains. CO-Net maintains the characteristics of high storage efficiency since models with the preponderance of shared parameters can be assembled into a single model. Specifically, we leverage residual MLP (Res-MLP) block for effective feature extraction and scale it gracefully along the depth and width of the network to meet the demands of different tasks. Based on the block, we propose a novel nested layer-wise processing policy, which identifies the optimal architecture for each task while provides partial sharing parameters and partial non-sharing parameters inside each layer of the block. Such policy tackles the inherent challenges of multi-task learning on point cloud, e.g., diverse model topologies resulting from task skew and conflicting gradients induced by heterogeneous dataset domains. Finally, we propose a sign-based gradient surgery to promote the training of CO-Net, thereby emphasizing the usage of task-shared parameters and guaranteeing that each task can be thoroughly optimized. Experimental results reveal that models optimized by CO-Net jointly for all point cloud tasks maintain much fewer computation cost and overall storage cost yet outpace prior methods by a significant margin. We also demonstrate that CO-Net allows incremental learning and prevents catastrophic amnesia when adapting to a new point cloud task.

Most nighttime semantic segmentation studies are based on domain adaptation approaches and image input. However, limited by the low dynamic range of conventional cameras, images fail to capture structural details and boundary information in low-light conditions. Event cameras, as a new form of vision sensors, are complementary to conventional cameras with their high dynamic range. To this end, we propose a novel unsupervised Cross-Modality Domain Adaptation (CMDA) framework to leverage multi-modality (Images and Events) information for nighttime semantic segmentation, with only labels on daytime images. In CMDA, we design the Image Motion-Extractor to extract motion information and the Image Content-Extractor to extract content information from images, in order to bridge the gap between different modalities (Images to Events) and domains (Day to Night). Besides, we introduce the first image-event nighttime semantic segmentation dataset. Extensive experiments on both the public image dataset and the proposed image-event dataset demonstrate the effectiveness of our proposed approach. We open-source our code, models, and dataset at https://github.com/XiaRho/CMDA.