CVPR 2022 Papers

In this paper, we propose a novel sequence verification task that aims to distinguish positive video pairs performing the same action sequence from negative ones with step-level transformations but still conducting the same task. Such a challenging task resides in an open-set setting without prior action detection or segmentation that requires event-level or even frame-level annotations. To that end, we carefully reorganize two publicly available action-related datasets with step-procedure-task structure. To fully investigate the effectiveness of any method, we collect a scripted video dataset enumerating all kinds of step-level transformations in chemical experiments. Besides, a novel evaluation metric Weighted Distance Ratio is introduced to ensure equivalence for different step-level transformations during evaluation. In the end, a simple but effective baseline based on the transformer encoder with a novel sequence alignment loss is introduced to better characterize long-term dependency between steps, which outperforms other action recognition methods. Codes and data will be released.

Although many video prediction methods have obtained good performance in low-resolution (64 128) videos, predictive models for high-resolution (512 4K) videos have not been fully explored yet, which are more meaningful due to the increasing demand for high-quality videos. Compared with low-resolution videos, high-resolution videos contain richer appearance (spatial) information and more complex motion (temporal) information. In this paper, we propose a Spatiotemporal Residual Predictive Model (STRPM) for high-resolution video prediction. On the one hand, we propose a Spatiotemporal Encoding-Decoding Scheme to preserve more spatiotemporal information for high-resolution videos. In this way, the appearance details for each frame can be greatly preserved. On the other hand, we design a Residual Predictive Memory (RPM) which focuses on modeling the spatiotemporal residual features (STRF) between previous and future frames instead of the whole frame, which can greatly help capture the complex motion information in high-resolution videos. In addition, the proposed RPM can supervise the spatial encoder and temporal encoder to extract different features in the spatial domain and the temporal domain, respectively. Moreover, the proposed model is trained using generative adversarial networks (GANs) with a learned perceptual loss (LP-loss) to improve the perceptual quality of the predictions. Experimental results show that STRPM can generate more satisfactory results compared with various existing methods.

Accurately detecting and tracking pedestrians in 3D space is challenging due to large variations in rotations, poses and scales. The situation becomes even worse for dense crowds with severe occlusions. However, existing benchmarks either only provide 2D annotations, or have limited 3D annotations with low-density pedestrian distribution, making it difficult to build a reliable pedestrian perception system especially in crowded scenes. To better evaluate pedestrian perception algorithms in crowded scenarios, we introduce a large-scale multimodal dataset, STCrowd. Specifically, in STCrowd, there are a total of 219K pedestrian instances and 20 persons per frame on average, with various levels of occlusion. We provide synchronized LiDAR point clouds and camera images as well as their corresponding 3D labels and joint IDs. STCrowd can be used for various tasks, including LiDAR-only, image-only, and sensor-fusion based pedestrian detection and tracking. We provide baselines for most of the tasks. In addition, considering the property of sparse global distribution and density-varying local distribution of pedestrians, we further propose a novel method, Density-aware Hierarchical heatmap Aggregation (DHA), to enhance pedestrian perception in crowded scenes. Extensive experiments show that our new method achieves state-of-the-art performance on the STCrowd dataset, especially on cases with severe occlusion. The dataset and code will be released to facilitate related research when the paper is published.

Video frame interpolation (VFI) is currently a very active research topic, with applications spanning computer vision, post production and video encoding. VFI can be extremely challenging, particularly in sequences containing large motions, occlusions or dynamic textures, where existing approaches fail to offer perceptually robust interpolation performance. In this context, we present a novel deep learning based VFI method, ST-MFNet, based on a Spatio-Temporal Multi-Flow architecture. ST-MFNet employs a new multi-scale multi-flow predictor to estimate many-to-one intermediate flows, which are combined with conventional one-to-one optical flows to capture both large and complex motions. In order to enhance interpolation performance for various textures, a 3D CNN is also employed to model the content dynamics over an extended temporal window. Moreover, ST-MFNet has been trained within an ST-GAN framework, which was originally developed for texture synthesis, with the aim of further improving perceptual interpolation quality. Our approach has been comprehensively evaluated -- compared with fourteen state-of-the-art VFI algorithms -- clearly demonstrating that ST-MFNet consistently outperforms these benchmarks on varied and representative test datasets, with significant gains up to 1.09dB in PSNR for cases including large motions and dynamic textures. Our source code is available at https://github.com/danielism97/ST-MFNet.

Self-training via pseudo labeling is a conventional, simple, and popular pipeline to leverage unlabeled data. In this work, we first construct a strong baseline of self-training (namely ST) for semi-supervised semantic segmentation via injecting strong data augmentations (SDA) on unlabeled images to alleviate overfitting noisy labels as well as decouple similar predictions between the teacher and student. With this simple mechanism, our ST outperforms all existing methods without any bells and whistles, e.g., iterative re-training. Inspired by the impressive results, we thoroughly investigate the SDA and provide some empirical analysis. Nevertheless, incorrect pseudo labels are still prone to accumulate and degrade the performance. To this end, we further propose an advanced self-training framework (namely ST++), that performs selective re-training via prioritizing reliable unlabeled images based on holistic prediction-level stability. Concretely, several model checkpoints are saved in the first stage supervised training, and the discrepancy of their predictions on the unlabeled image serves as a measurement for reliability. Our image-level selection offers holistic contextual information for learning. We demonstrate that it is more suitable for segmentation than common pixel-wise selection. As a result, ST++ further boosts the performance of our ST. Code is available at https://github.com/LiheYoung/ST-PlusPlus.

Conventional deep learning based methods for 3D object detection require a large amount of 3D bounding box annotations for training, which is expensive to obtain in practice. Sparsely annotated object detection, which can largely reduce the annotations, is very challenging since the missingannotated instances would be regarded as the background during training. In this paper, we propose a sparselysupervised 3D object detection method, named SS3D. Aiming to eliminate the negative supervision caused by the missing annotations, we design a missing-annotated instance mining module with strict filtering strategies to mine positive instances. In the meantime, we design a reliable background mining module and a point cloud filling data augmentation strategy to generate the confident data for iteratively learning with reliable supervision. The proposed SS3D is a general framework that can be used to learn any modern 3D object detector. Extensive experiments on the KITTI dataset reveal that on different 3D detectors, the proposed SS3D framework with only 20% annotations required can achieve on-par performance comparing to fullysupervised methods. Comparing with the state-of-the-art semi-supervised 3D objection detection on KITTI, our SS3D improves the benchmarks by significant margins under the same annotation workload. Moreover, our SS3D also outperforms the state-of-the-art weakly-supervised method by remarkable margins, highlighting its effectiveness.

Visual private information leakage is an emerging key issue for the fast growing applications of video understanding like activity recognition. Existing approaches for mitigating privacy leakage in action recognition require privacy labels along with the action labels from the video dataset. However, annotating frames of video dataset for privacy labels is not feasible. Recent developments of self-supervised learning (SSL) have unleashed the untapped potential of the unlabeled data. For the first time, we present a novel training framework which removes privacy information from input video in a self-supervised manner without requiring privacy labels. Our training framework consists of three main components: anonymization function, self-supervised privacy removal branch, and action recognition branch. We train our framework using a minimax optimization strategy to minimize the action recognition cost function and maximize the privacy cost function through a contrastive self-supervised loss. Employing existing protocols of known-action and privacy attributes, our framework achieves a competitive action-privacy trade-off to the existing state-of-the-art supervised methods. In addition, we introduce a new protocol to evaluate the generalization of learned the anonymization function to novel-action and privacy attributes and show that our self-supervised framework outperforms existing supervised methods. Code available at: https://github.com/DAVEISHAN/SPAct

Parametric 3D models have formed a fundamental role in modeling deformable objects, such as human bodies, faces, and hands; however, the construction of such parametric models requires significant manual intervention and domain expertise. Recently, neural implicit 3D representations have shown great expressibility in capturing 3D shape geometry. We observe that deformable object motion is often semantically structured, and thus propose to learn Structured-implicit PArametric Models (SPAMs) as a deformable object representation that structurally decomposes non-rigid object motion into part-based disentangled representations of shape and pose, with each being represented by deep implicit functions. This enables a structured characterization of object movement, with part decomposition characterizing a lower-dimensional space in which we can establish coarse motion correspondence. In particular, we can leverage the part decompositions at test time to fit to new depth sequences of unobserved shapes, by establishing part correspondences between the input observation and our learned part spaces; this guides a robust joint optimization between the shape and pose of all parts, even under dramatic motion sequences. Experiments demonstrate that our part-aware shape and pose understanding lead to state-of-the-art performance in reconstruction and tracking of depth sequences of complex deforming object motion.

Spherical novel view synthesis (SNVS) is the task of estimating 360 views at dynamic novel views given a set of 360 input views. Prior arts learn multi-sphere image (MSI) representations that enables fast rendering times but are only limited to modelling low-dimensional color values. Modelling high-dimensional appearance features in MSI can result in better view synthesis, but it is not feasible to represent high-dimensional features in a large number (>64) of MSI spheres. We propose a novel MSI representation called Soft Occlusion MSI (SOMSI) that enables modelling high-dimensional appearance features in MSI while retaining the fast rendering times of a standard MSI. Our key insight is to model appearance features in a smaller set (e.g. 3) of occlusion levels instead of larger number of MSI levels. Experiments on both synthetic and real-world scenes demonstrate that using SOMSI can provide a good balance between accuracy and runtime. SOMSI can produce considerably better results compared to MSI based MODS, while having similar fast rendering time. SOMSI view synthesis quality is on-par with state-of-the-art NeRF like model while being 2 orders of magnitude faster. For code, additional results and data, please visit https://tedyhabtegebrial.github.io/somsi.

We present a self-supervised method to learn dynamic 3D deformations of garments worn by parametric human bodies. State-of-the-art data-driven approaches to model 3D garment deformations are trained using supervised strategies that require large datasets, usually obtained by expensive physics-based simulation methods or professional multi-camera capture setups. In contrast, we propose a new training scheme that removes the need for ground-truth samples, enabling self-supervised training of dynamic 3D garment deformations. Our key contribution is to realize that physics-based deformation models, traditionally solved in a frame-by-frame basis by implicit integrators, can be recasted as an optimization problem. We leverage such optimization-based scheme to formulate a set of physics-based loss terms that can be used to train neural networks without precomputing ground-truth data. This allows us to learn models for interactive garments, including dynamic deformations and fine wrinkles, with two orders of magnitude speed up in training time compared to state-of-the-art supervised methods.

This paper presents a new solution for low-light image enhancement by collectively exploiting Signal-to-Noise-Ratio-aware transformers and convolutional models to dynamically enhance pixels with spatial-varying operations. They are long-range operations for image regions of extremely low Signal-to-Noise-Ratio (SNR) and short-range operations for other regions. We propose to take an SNR prior to guide the feature fusion and formulate the SNR-aware transformer with a new self-attention model to avoid tokens from noisy image regions of very low SNR. Extensive experiments show that our framework consistently achieves better performance than SOTA approaches on seven representative benchmarks with the same structure. Also, we conducted a large-scale user study with 100 participants to verify the superior perceptual quality of our results.

A variety of diagnostic and therapeutic protocols rely on locating in vivo target anatomical structures, which can be obtained from medical scans. However, organs move and deform as the patient changes his/her pose. In order to obtain accurate target location information, clinicians have to either conduct frequent intraoperative scans, resulting in higher exposition of patients to radiations, or adopt proxy procedures (e.g., creating and using custom molds to keep patients in the exact same pose during both preoperative organ scanning and subsequent treatment. Such custom proxy methods are typically sub-optimal, constraining the clinicians and costing precious time and money to the patients. To the best of our knowledge, this work is the first to present a learning-based approach to estimate the patient's internal organ deformation for arbitrary human poses in order to assist with radiotherapy and similar medical protocols. The underlying method first leverages medical scans to learn a patient-specific representation that potentially encodes the organ's shape and elastic properties. During inference, given the patient's current body pose information and the organ's representation extracted from previous medical scans, our method can estimate their current organ deformation to offer guidance to clinicians. We conduct experiments on a well-sized dataset which is augmented through real clinical data using finite element modeling. Our results suggest that pose-dependent organ deformation can be learned through a point cloud autoencoder conditioned on the parametric pose input. We hope that this work can be a starting point for future research towards closing the loop between human mesh recovery and anatomical reconstruction, with applications beyond the medical domain.

Self-supervised methods have significantly closed the gap with end-to-end supervised learning for image classification [13,24]. In the case of human action videos, however, where both appearance and motion are significant factors of variation, this gap remains significant [28,58]. One of the key reasons for this is that sampling pairs of similar video clips, a required step for many self-supervised contrastive learning methods, is currently done conservatively to avoid false positives. A typical assumption is that similar clips only occur temporally close within a single video, leading to insufficient examples of motion similarity. To mitigate this, we propose SLIC, a clustering-based self-supervised contrastive learning method for human action videos. Our key contribution is that we improve upon the traditional intra-video positive sampling by using iterative clustering to group similar video instances. This enables our method to leverage pseudo-labels from the cluster assignments to sample harder positives and negatives. SLIC outperforms state-of-the-art video retrieval baselines by +15.4% on top-1 recall on UCF101 and by +5.7% when directly transferred to HMDB51. With end-to-end finetuning for action classification, SLIC achieves 83.2% top-1 accuracy (+0.8%) on UCF101 and 54.5% on HMDB51 (+1.6%). SLIC is also competitive with the state-of-the-art in action classification after self-supervised pretraining on Kinetics400.

Object detection under imperfect data receives great attention recently. Weakly supervised object detection (WSOD) suffers from severe localization issues due to the lack of instance-level annotation, while semi-supervised object detection (SSOD) remains challenging led by the inter-image discrepancy between labeled and unlabeled data. In this study, we propose the Single Instance annotated Object Detection (SIOD), requiring only one instance annotation for each existing category in an image. Degraded from inter-task (WSOD) or inter-image (SSOD) discrepancies to the intra-image discrepancy, SIOD provides more reliable and rich prior knowledge for mining the rest of unlabeled instances and trades off the annotation cost and performance. Under the SIOD setting, we propose a simple yet effective framework, termed Dual-Mining (DMiner), which consists of a Similarity-based Pseudo Label Generating module (SPLG) and a Pixel-level Group Contrastive Learning module (PGCL). SPLG firstly mines latent instances from feature representation space to alleviate the annotation missing problem. To avoid being misled by inaccurate pseudo labels, we propose PGCL to boost the tolerance to false pseudo labels. Extensive experiments on MS COCO verify the feasibility of the SIOD setting and the superiority of the proposed method, which obtains consistent and significant improvements compared to baseline methods and achieves comparable results with fully supervised object detection (FSOD) methods with only 40% instances annotated.

Real-world autonomous driving datasets comprise of images aggregated from different drives on the road. The ability to relight captured scenes to unseen lighting conditions, in a controllable manner, presents an opportunity to augment datasets with a richer variety of lighting conditions, similar to what would be encountered in the real-world. This paper presents a novel image-based relighting pipeline, SIMBAR, that can work with a single image as input. To the best of our knowledge, there is no prior work on scene relighting leveraging explicit geometric representations from a single image. We present qualitative comparisons with prior multi-view scene relighting baselines. To further validate and effectively quantify the benefit of leveraging SIMBAR for data augmentation for automated driving vision tasks, object detection and tracking experiments are conducted with a state-of-the-art method, a Multiple Object Tracking Accuracy (MOTA) of 93.3% is achieved with CenterTrack on SIMBAR-augmented KITTI - an impressive 9.0% relative improvement over the baseline MOTA of 85.6% with CenterTrack on original KITTI, both models trained from scratch and tested on Virtual KITTI. For more details and SIMBAR relit datasets, please visit our project website (https://simbarv1.github.io/).

Domain Adaptive Object Detection (DAOD) leverages a labeled domain to learn an object detector generalizing to a novel domain free of annotations. Recent advances align class-conditional distributions by narrowing down cross-domain prototypes (class centers). Though great success, they ignore the significant within-class variance and the domain-mismatched semantics within the training batch, leading to a sub-optimal adaptation. To overcome these challenges, we propose a novel SemantIc-complete Graph MAtching (SIGMA) framework for DAOD, which completes mismatched semantics and reformulates the adaptation with graph matching. Specifically, we design a Graph-embedded Semantic Completion module (GSC) that completes mismatched semantics through generating hallucination graph nodes in missing categories. Then, we establish cross-image graphs to model class-conditional distributions and learn a graph-guided memory bank for better semantic completion in turn. After representing the source and target data as graphs, we reformulate the adaptation as a graph matching problem, i.e., finding well-matched node pairs across graphs to reduce the domain gap, which is solved with a novel Bipartite Graph Matching adaptor (BGM). In a nutshell, we utilize graph nodes to establish semantic-aware node affinity and leverage graph edges as quadratic constraints in a structure-aware matching loss, achieving fine-grained adaptation with a node-to-node graph matching. Extensive experiments verify that SIGMA outperforms existing works significantly. Our codes are available at https://github.com/CityU-AIM-Group/SIGMA.

Adapting to a continuously evolving environment is a safety-critical challenge inevitably faced by all autonomous-driving systems. Existing image- and video-based driving datasets, however, fall short of capturing the mutable nature of the real world. In this paper, we introduce the largest synthetic dataset for autonomous driving, SHIFT. It presents discrete and continuous shifts in cloudiness, rain and fog intensity, time of day, and vehicle and pedestrian density. Featuring a comprehensive sensor suite and annotations for several mainstream perception tasks, SHIFT allows to investigate how a perception systems' performance degrades at increasing levels of domain shift, fostering the development of continuous adaptation strategies to mitigate this problem and assessing the robustness and generality of a model. Our dataset and benchmark toolkit are publicly available at https://www.vis.xyz/shift.

Scene Graph Generation (SGG) remains a challenging visual understanding task due to its compositional property. Most previous works adopt a bottom-up two-stage or a point-based one-stage approach, which often suffers from high time complexity or sub-optimal designs. In this work, we propose a novel SGG method to address the aforementioned issues, formulating the task as a bipartite graph construction problem. To solve the problem, we develop a transformer-based end-to-end framework that first generates the entity and predicate proposal set, followed by inferring directed edges to form the relation triplets. In particular, we develop a new entity-aware predicate representation based on a structural predicate generator that leverages the compositional property of relationships. Moreover, we design a graph assembling module to infer the connectivity of the bipartite scene graph based on our entity-aware structure, enabling us to generate the scene graph in an end-to-end manner. Extensive experimental results show that our design is able to achieve the state-of-the-art or comparable performance on two challenging benchmarks, surpassing most of the existing approaches and enjoying higher efficiency in inference. We hope our model can serve as a strong baseline for the Transformer-based scene graph generation. Code is available in https://github.com/Scarecrow0/SGTR

Talking gesture generation is a practical yet challenging task which aims to synthesize gestures in line with speech. Gestures with meaningful signs can better convey useful information and arouse sympathy in the audience. Current works focus on aligning gestures with the speech rhythms, which are hard to mine the semantics and model semantic gestures explicitly. In this paper, we propose a novel method SEmantic Energized Generation (SEEG), for semantic-aware gesture generation. Our method contains two parts: DEcoupled Mining module (DEM) and Semantic Energizing Module (SEM). DEM decouples the semantic-irrelevant information from inputs and separately mines information for the beat and semantic gestures. SEM conducts semantic learning and produces semantic gestures. Apart from representational similarity, SEM requires the predictions to express the same semantics as the ground truth. Besides, a semantic prompter is designed in SEM to leverage the semantic-aware supervision to predictions. This promotes the networks to learn and generate semantic gestures. Experimental results reported in three metrics on different benchmarks prove that SEEG efficiently mines semantic cues and generates semantic gestures. In comparison, SEEG outperforms other methods in all semantic-aware evaluations on different datasets. Qualitative evaluations also indicate the superiority of SEEG in semantic expressiveness.