CVPR 2021 Papers

In this paper, we tackle the problem of dynamic scene deblurring. Most existing deep end-to-end learning approaches adopt the same generic model for all unseen test images. These solutions are sub-optimal, as they fail to utilize the internal information within a specific image. On the other hand, a self-supervised approach, SelfDeblur, enables internal-training within a test image from scratch, but it does not fully take advantages of large external dataset. In this work, we propose a novel self-supervised meta-auxiliary learning to improve the performance of deblurring by integrating both external and internal learning. Concretely, we build a self-supervised auxiliary reconstruction task which shares a portion of the network with the primary deblurring task. The two tasks are jointly trained on an external dataset. Furthermore, we propose a meta-auxiliary training scheme to further optimize the pre-trained model as a base learner which is applicable for fast adaptation at test time. During training, the performance of both tasks is coupled. Therefore, we are able to exploit the internal information at test time via the auxiliary task to enhance the performance of deblurring. Extensive experimental results across evaluation datasets demonstrate the effectiveness of test-time adaptation of the proposed method.

We consider the task of 3D pose estimation and trackingof multiple people seen in an arbitrary number of camerafeeds. We propose TesseTrack, a novel top-down approachthat simultaneously reasons about multiple individuals' 3Dbody joint reconstructions and associations in space andtime in a single end-to-end learnable framework. At the core of our approach is a novel spatio-temporal formulation that operates in a common voxelized feature space aggregated from single- or multiple-camera views. After a person detection step, a 4D CNN produces short-term person-specific representations which are then linked across time by a differentiable matcher. The linked descriptions are then merged and deconvolved into 3D poses. This joint spatio-temporal formulation contrasts with previous piece-wise strategies that treat 2D pose estimation, 2D-to-3D lifting, and 3D pose tracking as independent sub-problems that are error-prone when solved in isolation. Furthermore, unlike previous methods, TesseTrack is robust to changes in the number of camera views and achieves very good results even if a single view is available at inference time. Quantitative evaluation of 3D pose reconstruction accuracy on standard benchmarks shows significant improvements over the state of the art. Evaluation of multi-person articulated 3D pose tracking in our novel evaluation framework demonstrates the superiority of TesseTrack over strong baselines.

Action segmentation refers to inferring boundaries of semantically consistent visual concepts in videos and is an important requirement for many video understanding tasks. For this and other video understanding tasks, supervised approaches have achieved encouraging performance but require a high volume of detailed, frame-level, annotations. We present a fully automatic and unsupervised approach for segmenting actions in a video that does not require any training. Our proposal is an effective temporally-weighted hierarchical clustering algorithm that can group semantically consistent frames of the video. The main finding is that representing a video with a 1-nearest neighbor graph by taking into account the time progression is sufficient to form semantically and temporally consistent clusters of frames where each cluster may represent some action in the video. Additionally, we establish strong unsupervised baselines for action segmentation and show significant performance improvements over published unsupervised methods on five challenging action segmentation datasets. Our code is available at https://github.com/ssarfraz/FINCH-Clustering/tree/master/TW-FINCH

We propose a novel approach to few-shot action recognition, finding temporally-corresponding frame tuples between the query and videos in the support set. Distinct from previous few-shot works, we construct class prototypes using the CrossTransformer attention mechanism to observe relevant sub-sequences of all support videos, rather than using class averages or single best matches. Video representations are formed from ordered tuples of varying numbers of frames, which allows sub-sequences of actions at different speeds and temporal offsets to be compared. Our proposed Temporal-Relational CrossTransformers (TRX) achieve state-of-the-art results on few-shot splits of Kinetics, Something-Something V2 (SSv2), HMDB51 and UCF101. Importantly, our method outperforms prior work on SSv2 by a wide margin (12%) due to the its ability to model temporal relations. A detailed ablation showcases the importance of matching to multiple support set videos and learning higher-order relational CrossTransformers.

Our objective in this work is fine-grained classification of actions in untrimmed videos, where the actions may be temporally extended or may span only a few frames of the video. We cast this into a query-response mechanism, where each query addresses a particular question, and has its own response label set. We make the following four contributions: (i) We propose a new model---a Temporal Query Network---which enables the query-response functionality, and a structural understanding of fine-grained actions. It attends to relevant segments for each query with a temporal attention mechanism, and can be trained using only the labels for each query. (ii) We propose a new way---stochastic feature bank update---to train a network on videos of various lengths with the dense sampling required to respond to fine-grained queries. (iii) we compare the TQN to other architectures and text supervision methods, and analyze their pros and cons. Finally, (iv) we evaluate the method extensively on the FineGym and Diving48 benchmarks for fine-grained action classification and surpass the state-of-the-art using only RGB features.

Space-time video super-resolution (STVSR) aims to increase the spatial and temporal resolutions of low-resolution and low-frame-rate videos. Recently, deformable convolution based methods have achieved promising STVSR performance, but they could only infer the intermediate frame pre-defined in the training stage. Besides, these methods undervalued the short-term motion cues among adjacent frames. In this paper, we propose a Temporal Modulation Network (TMNet) to interpolate arbitrary intermediate frame(s) with accurate high-resolution reconstruction. Specifically, we propose a Temporal Modulation Block (TMB) to modulate deformable convolution kernels for controllable feature interpolation. To well exploit the temporal information, we propose a Locally-temporal Feature Comparison (LFC) module, along with the Bi-directional Deformable ConvLSTM, to extract short-term and long-term motion cues in videos. Experiments on three benchmark datasets demonstrate that our TMNet outperforms previous STVSR methods. The code is available at https://github.com/CS-GangXu/TMNet.

Temporal action proposal generation aims to estimate temporal intervals of actions in untrimmed videos, which is a challenging yet important task in the video understanding field. The proposals generated by current methods still suffer from inaccurate temporal boundaries and inferior confidence used for retrieval owing to the lack of efficient temporal modeling and effective boundary context utilization. In this paper, we propose Temporal Context Aggregation Network (TCANet) to generate high-quality action proposals through local and global temporal context aggregation and complementary as well as progressive boundary refinement. Specifically, we first design a Local-Global Temporal Encoder (LGTE), which adopts the channel grouping strategy to efficiently encode both local and global temporal inter-dependencies. Furthermore, both the boundary and internal context of proposals are adopted for frame-level and segment-level boundary regressions, respectively. Temporal Boundary Regressor (TBR) is designed to combine these two regression granularities in an end-to-end fashion, which achieves the precise boundaries and reliable confidence of proposals through progressive refinement. Extensive experiments are conducted on three challenging datasets: HACS, ActivityNet-v1.3, and THUMOS-14, where TCANet can generate proposals with high precision and recall. By combining with the existing action classifier, TCANet can obtain remarkable temporal action detection performance compared with other methods. Not surprisingly, the proposed TCANet won the 1st place in the CVPR 2020 - HACS challenge leaderboard on temporal action localization task.

Temporal action segmentation approaches have been very successful recently. However, annotating videos with frame-wise labels to train such models is very expensive and time consuming. While weakly supervised methods trained using only ordered action lists require less annotation effort, the performance is still worse than fully supervised approaches. In this paper, we propose to use timestamp supervision for the temporal action segmentation task. Timestamps require a comparable annotation effort to weakly supervised approaches, and yet provide a more supervisory signal. To demonstrate the effectiveness of timestamp supervision, we propose an approach to train a segmentation model using only timestamps annotations. Our approach uses the model output and the annotated timestamps to generate frame-wise labels by detecting the action changes. We further introduce a confidence loss that forces the predicted probabilities to monotonically decrease as the distance to the timestamps increases. This ensures that all and not only the most distinctive frames of an action are learned during training. The evaluation on four datasets shows that models trained with timestamps annotations achieve comparable performance to the fully supervised approaches.

In this work, we propose TediGAN, a novel framework for multi-modal image generation and manipulation with textual descriptions. The proposed method consists of three components: StyleGAN inversion module, visual-linguistic similarity learning, and instance-level optimization. The inversion module maps real images to the latent space of a well-trained StyleGAN. The visual-linguistic similarity learns the text-image matching by mapping the image and text into a common embedding space. The instance-level optimization is for identity preservation in manipulation. Our model can produce diverse and high-quality images with an unprecedented resolution at 1024 x 1024. Using a control mechanism based on style-mixing, our TediGAN inherently supports image synthesis with multi-modal inputs, such as sketches or semantic labels, with or without instance guidance. To facilitate text-guided multi-modal synthesis, we propose the Multi-Modal CelebA-HQ, a large-scale dataset consisting of real face images and corresponding semantic segmentation map, sketch, and textual descriptions. Extensive experiments on the introduced dataset demonstrate the superior performance of our proposed method. Code and data are available at https://github.com/weihaox/TediGAN.

Topology matters. Despite the recent success of point cloud processing with geometric deep learning, it remains arduous to capture the complex topologies of point cloud data with a learning model. Given a point cloud dataset containing objects with various genera, or scenes with multiple objects, we propose an autoencoder, TearingNet, which tackles the challenging task of representing the point clouds using a fixed-length descriptor. Unlike existing works directly deforming predefined primitives of genus zero (e.g., a 2D square patch) to an object-level point cloud, our TearingNet is characterized by a proposed Tearing network module and a Folding network module interacting with each other iteratively. Particularly, the Tearing network module learns the point cloud topology explicitly. By breaking the edges of a primitive graph, it tears the graph into patches or with holes to emulate the topology of a target point cloud, leading to faithful reconstructions. Experimentation shows the superiority of our proposal in terms of reconstructing point clouds as well as generating more topology-friendly representations than benchmarks.

Generative Adversarial Networks (GANs) have achieved huge success in generating high-fidelity images, however, they suffer from low efficiency due to tremendous computational cost and bulky memory usage. Recent efforts on compression GANs show noticeable progress in obtaining smaller generators by sacrificing image quality or involving a time-consuming searching process. In this work, we aim to address these issues by introducing a teacher network that provides a search space in which efficient network architectures can be found, in addition to performing knowledge distillation. First, we revisit the search space of generative models, introducing an inception-based residual block into generators. Second, to achieve target computation cost, we propose a one-step pruning algorithm that searches a student architecture from the teacher model and substantially reduces searching cost. It requires no L1 sparsity regularization and its associated hyper-parameters, simplifying the training procedure. Finally, we propose to distill knowledge through maximizing feature similarity between teacher and student via an index named Global Centered Kernel Alignment (GCKA). Our compressed networks achieve better image fidelity (FID, mIoU) than the original models with much-reduced computational cost, e.g., MACs.

Many computer vision tasks address the problem of scene understanding and are naturally interrelated e.g. object classification, detection, scene segmentation, depth estimation, etc. We show that we can leverage the inherent relationships among collections of tasks, as they are trained jointly, supervising each other through their known relationships via consistency losses. Furthermore, explicitly utilizing the relationships between tasks allows improving their performance while dramatically reducing the need for labeled data, and allows training with additional unsupervised or simulated data. We demonstrate a distributed joint training algorithm with task-level parallelism, which affords a high degree of asynchronicity and robustness. This allows learning across multiple tasks, or with large amounts of input data, at scale. We demonstrate our framework on subsets of the following collection of tasks: depth and normal prediction, semantic segmentation, 3D motion and ego-motion estimation, and object tracking and 3D detection in point clouds. We observe improved performance across these tasks, especially in the low-label regime.

Often, labeling large amount of data is challenging due to high labeling cost limiting the application domain of deep learning techniques. Active learning (AL) tackles this by querying the most informative samples to be annotated among unlabeled pool. Two promising directions for AL that have been recently explored are task-agnostic approach to select data points that are far from the current labeled pool and task-aware approach that relies on the perspective of task model. Unfortunately, the former does not exploit structures from tasks and the latter does not seem to well-utilize overall data distribution. Here, we propose task-aware variational adversarial AL (TA-VAAL) that modifies task-agnostic VAAL, that considered data distribution of both label and unlabeled pools, by relaxing task learning loss prediction to ranking loss prediction and by using ranking conditional generative adversarial network to embed normalized ranking loss information on VAAL. Our proposed TA-VAAL outperforms state-of-the-arts on various benchmark datasets for classifications with balanced / imbalanced labels as well as semantic segmentation and its task-aware and task-agnostic AL properties were confirmed with our in-depth analyses.

Specialized domain knowledge is often necessary to accurately annotate training sets for in-depth analysis, but can be burdensome and time-consuming to acquire from domain experts. This issue arises prominently in automated behavior analysis, in which agent movements or actions of interest are detected from video tracking data. To reduce annotation effort, we present TREBA: a method to learn annotation-sample efficient trajectory embedding for behavior analysis, based on multi-task self-supervised learning. The tasks in our method can be efficiently engineered by domain experts through a process we call "task programming", which uses programs to explicitly encode structured knowledge from domain experts. Total domain expert effort can be reduced by exchanging data annotation time for the construction of a small number of programmed tasks. We evaluate this trade-off using data from behavioral neuroscience, in which specialized domain knowledge is used to identify behaviors. We present experimental results in three datasets across two domains: mice and fruit flies. Using embeddings from TREBA, we reduce annotation burden by up to a factor of 10 without compromising accuracy compared to state-of-the-art features. Our results thus suggest that task programming and self-supervision can be an effective way to reduce annotation effort for domain experts.

This paper addresses the task of unsupervised video multi-object segmentation. Current approaches follow a two-stage paradigm: 1) detect object proposals using pre-trained Mask R-CNN, and 2) conduct generic feature matching for temporal association using re-identification techniques. However, the generic features, widely used in both stages, are not reliable for characterizing unseen objects, leading to poor generalization. To address this, we introduce a novel approach for more accurate and efficient spatio-temporal segmentation. In particular, to address instance discrimination, we propose to combine foreground region estimation and instance grouping together in one network, and additionally introduce temporal guidance for segmenting each frame, enabling more accurate object discovery. For temporal association, we complement current video object segmentation architectures with a discriminative appearance model, capable of capturing more fine-grained target-specific information. Given object proposals from the instance discrimination network, three essential strategies are adopted to achieve accurate segmentation: 1) target-specific tracking using a memory-augmented appearance model; 2) target-agnostic verification to trace possible tracklets for the proposal; 3) adaptive memory updating using the verified segments. We evaluate the proposed approach on DAVIS_ 17 and YouTube-VIS, and the results demonstrate that it outperforms state-of-the-art methods both in segmentation accuracy and inference speed.

We address the problem of performing backpropagation for computation graphs involving 3D transformation groups SO(3), SE(3), and Sim(3). 3D transformation groups are widely used in 3D vision and robotics, but they do not form vector spaces and instead lie on smooth manifolds. The standard backpropagation approach, which embeds 3D transformations in Euclidean spaces, suffers from numerical difficulties. We introduce a new library, which exploits the group structure of 3D transformations and performs backpropagation in the tangent spaces of manifolds. We show that our approach is numerically more stable, easier to implement, and beneficial to a diverse set of tasks. Our plug-and-play PyTorch library is available at https://github.com/princeton-vl/lietorch.

Designed to learn long-range interactions on sequential data, transformers continue to show state-of-the-art results on a wide variety of tasks. In contrast to CNNs, they contain no inductive bias that prioritizes local interactions. This makes them expressive, but also computationally infeasible for long sequences, such as high-resolution images. We demonstrate how combining the effectiveness of the inductive bias of CNNs with the expressivity of transformers enables them to model and thereby synthesize high-resolution images. We show how to (i) use CNNs to learn a context-rich vocabulary of image constituents, and in turn (ii) utilize transformers to efficiently model their composition within high-resolution images. Our approach is readily applied to conditional synthesis tasks, where both non-spatial information, such as object classes, and spatial information, such as segmentations, can control the generated image. In particular, we present the first results on semantically-guided synthesis of megapixel images with transformers.

By the one-to-many nature of the super-resolution (SR) problem, a single low-resolution (LR) image can be mapped to many high-resolution (HR) images. However, learning based SR algorithms are trained to map an LR image to the corresponding ground truth (GT) HR image in the training dataset. The training loss will increase and penalize the algorithm when the output does not exactly match the GT target, even when the outputs are mathematically valid candidates according to the SR framework. This becomes more problematic for the blind SR, as diverse unknown blur kernels exacerbate the ill-posedness of the problem. To this end, we propose a fundamentally different approach for the SR by introducing the concept of the adaptive target. The adaptive target is generated from the original GT target by a transformation to match the output of the SR network. The adaptive target provides an effective way for the SR algorithm to deal with the ill-posed nature of the SR, by providing the algorithm with the flexibility of accepting a variety of valid solutions. Experimental results show the effectiveness of our algorithm, especially for improving the perceptual quality of HR outputs.

The ability to segment teeth precisely from digitized 3D dental models is an essential task in computer-aided orthodontic surgical planning. To date, deep learning based methods have been popularly used to handle this task. State-of-the-art methods directly concatenate the raw attributes of 3D inputs, namely coordinates and normal vectors of mesh cells, to train a single-stream network for fully-automated tooth segmentation. This, however, has the drawback of ignoring the different geometric meanings provided by those raw attributes. This issue might possibly confuse the network in learning discriminative geometric features and result in many isolated false predictions on the dental model. Against this issue, we propose a two-stream graph convolutional network (TSGCNet) to learn multi-view geometric information from different geometric attributes. Our TSGCNet adopts two graph-learning streams, designed in an input-aware fashion, to extract more discriminative high-level geometric representations from coordinates and normal vectors, respectively. These feature representations learned from the designed two different streams are further fused to integrate the multi-view complementary information for the cell-wise dense prediction task. We evaluate our proposed TSGCNet on a real-patient dataset of dental models acquired by 3D intraoral scanners, and experimental results demonstrate that our method significantly outperforms state-of-the-art methods for 3D shape segmentation.