CVPR 2021 Papers

We propose a new framework for understanding and representing related salient events in a video using visual semantic role labeling. We represent videos as a set of related events, wherein each event consists of a verb and multiple entities that fulfill various roles relevant to that event. To study the challenging task of semantic role labeling in videos or VidSRL, we introduce the VidSitu benchmark, a large scale video understanding data source with 27K 10-second movie clips richly annotated with a verb and semantic-roles every 2 seconds. Entities are co-referenced across events within a movie clip and events are connected to each other via event-event relations. Clips in VidSitu are drawn from a large collection of movies ( 3K) and have been chosen to be both complex ( 4.2 unique verbs within a video) as well as diverse ( 200 verbs have more than 100 annotations each). We provide a comprehensive analysis of the dataset in comparison to other publicly available video understanding benchmarks, several illustrative baselines and evaluate a range of standard video recognition models. Our code and dataset will be released publicly.

There has been a significant recent progress in the field of Embodied AI with researchers developing models and algorithms enabling embodied agents to navigate and interact within completely unseen environments. In this paper, we propose a new dataset and baseline models for the task of Rearrangement. We particularly focus on the task of Room Rearrangement: an agent begins by exploring a room and recording objects' initial configurations. We then remove the agent and change the poses and states (e.g., open/closed) of some objects in the room. The agent must restore the initial configurations of all objects in the room. Our dataset, named RoomR, includes 6,000 distinct rearrangement settings involving 72 different object types in 120 scenes. Our experiments show that solving this challenging interactive task that involves navigation and object interaction is beyond the capabilities of the current state-of-the-art techniques for embodied tasks and we are still very far from achieving perfect performance on these types of tasks.

This work focuses on object goal visual navigation, aiming at finding the location of an object from a given class, where in each step the agent is provided with an egocentric RGB image of the scene. We propose to learn the agent's policy using a reinforcement learning algorithm. Our key contribution is a novel attention probability model for visual navigation tasks. This attention encodes semantic information about observed objects, as well as spatial information about their place. This combination of the "what"" and the "where"" allows the agent to navigate toward the sought-after object effectively. The attention model is shown to improve the agent's policy and to achieve state-of-the-art results on commonly-used datasets.

Recently, deep face recognition has achieved significant progress because of Convolutional Neural Networks (CNNs) and large-scale datasets. However, training CNNs on a large-scale face recognition dataset with limited computational resources is still a challenge. This is because the classification paradigm needs to train a fully connected layer as the category classifier, and its parameters will be in the hundreds of millions if the training dataset contains millions of identities. This requires many computational resources, such as GPU memory. The metric learning paradigm is an economical computation method, but its performance is greatly inferior to that of the classification paradigm. To address this challenge, we propose a simple but effective CNN layer called the Virtual fully connected (Virtual FC) layer to reduce the computational consumption of the classification paradigm. Without bells and whistles, the proposed Virtual FC reduces the parameters by more than 100 times with respect to the fully connected layer and achieves competitive performance on mainstream face recognition evaluation datasets. Moreover, the performance of our Virtual FC layer on the evaluation datasets is superior to that of the metric learning paradigm by a significant margin. Our code will be released in hopes of disseminating our idea to other domains.

The de-facto approach to many vision tasks is to start from pretrained visual representations, typically learned via supervised training on ImageNet. Recent methods have explored unsupervised pretraining to scale to vast quantities of unlabeled images. In contrast, we aim to learn high-quality visual representations from fewer images. To this end we revisit supervised pretraining, and seek data-efficient alternatives to classification-based pretraining. We propose VirTex -- a pretraining approach using semantically dense captions to learn visual representations. We train convolutional networks from scratch on COCO Captions, and transfer them to downstream recognition tasks including image classification, object detection, and instance segmentation. On all tasks, VirTex yields features that match or exceed those learned on ImageNet -- supervised or unsupervised -- despite using up to ten times fewer images.

Recently, exploiting the effect of the unlabeled data for face recognition attracts increasing attention. However, there are still few works considering the situation that the unlabeled data is shallow which widely exists in real-world scenarios. The existing semi-supervised face recognition methods which focus on generating pseudo labels or minimizing softmax classification probabilities of the unlabeled data don't work very well on the unlabeled shallow data. Thus, it is still a challenge on how to effectively utilize the unlabeled shallow face data for improving the performance of face recognition. In this paper, we propose a novel face recognition method, named VirFace, to effectively apply the unlabeled shallow data for face recognition. VirFace consists of VirClass and VirInstance. Specifically, VirClass enlarges the inter-class distance by injecting the unlabeled data as new identities. Furthermore, VirInstance produces virtual instances sampled from the learned distribution of each identity to further enlarge the inter-class distance. To the best of our knowledge, we are the first working on tackling the unlabeled shallow face data. Extensive experiments have been conducted on both the small- and large-scale datasets, e.g. LFW and IJB-C, etc, showing the superiority of the proposed method.

This paper presents a detailed study of improving vision features and develops an improved object detection model for vision language (VL) tasks. Compared to the most widely used bottom-up and top-down model [2], the new model is bigger, pre-trained on much larger training corpora that combine multiple public annotated object detection datasets, and thus can generate representations of a richer collection of visual objects and concepts. While previous VL research focuses solely on improving the vision-language fusion model and leaves the object detection model improvement untouched, we present an empirical study to show that vision features matter significantly in VL models. In our experiments we feed the vision features generated by the new object detection model into a pre-trained transformer-based VL fusion model Oscar+, and fine-tune Oscar+ on a wide range of downstream VL tasks. Our results show that the new vision features significantly improve the performance across all VL tasks, creating new state-of-the-art results on seven public benchmarks. We will release the new object detection model to public.

This paper presents a view-guided solution for the task of point cloud completion. Unlike most existing methods directly inferring the missing points using shape priors, we address this task by introducing ViPC (view-guided point cloud completion) that takes the missing crucial global structure information from an extra single-view image. By leveraging a framework which sequentially performs effective cross-modality and cross-level fusions, our method achieves significantly superior results over typical existing solutions on a new large-scale dataset we collect for the view-guided point cloud completion task.

Humans can easily infer the underlying 3D geometry and texture of an object only from a single 2D image. Current computer vision methods can do this, too, but suffer from view generalization problems -- the models inferred tend to make poor predictions of appearance in novel views. As for generalization problems in machine learning, the difficulty is balancing single-view accuracy (cf. training error; bias) with novel view accuracy (cf. test error; variance). We describe a class of models whose geometric rigidity is easily controlled to manage this tradeoff. We describe a cycle consistency loss that improves view generalization (roughly, a model from a generated view should predict the original view well). View generalization of textures requires that models share texture information, so a car seen from the back still has headlights because other cars have headlights. We describe a cycle consistency loss that encourages model textures to be aligned, so as to encourage sharing. We compare our method against the state-of-the-art method and show both qualitative and quantitative improvements.

MoCo is effective for unsupervised image representation learning. In this paper, we propose VideoMoCo for unsupervised video representation learning. Given a video sequence as an input sample, we improve the temporal feature representations of MoCo from two perspectives. First, we introduce a generator to drop out several frames from this sample temporally. The discriminator is then learned to encode similar feature representations regardless of frame removals. By adaptively dropping out different frames during training iterations of adversarial learning, we augment this input sample to train a temporally robust encoder. Second, we use temporal decay to model key attenuation in the memory queue when computing the contrastive loss. As the momentum encoder updates after keys enqueue, the representation ability of these keys degrades when we use the current input sample for contrastive learning. This degradation is reflected via temporal decay to attend the input sample to recent keys in the queue. As a result, we adapt MoCo to learn video representations without empirically designing pretext tasks. By empowering the temporal robustness of the encoder and modeling the temporal decay of the keys, our VideoMoCo improves MoCo temporally based on contrastive learning. Experiments on benchmark datasets including UCF101 and HMDB51 show that VideoMoCo stands as a state-of-the-art video representation learning method.

This paper addresses the video rescaling task, which arises from the needs of adapting the video spatial resolution to suit individual viewing devices. We aim to jointly optimize video downscaling and upscaling as a combined task. Most recent studies focus on image-based solutions, which do not consider temporal information. We present two joint optimization approaches based on invertible neural networks with coupling layers. Our Long Short-Term Memory Video Rescaling Network (LSTM-VRN) leverages temporal information in the low-resolution video to form an explicit prediction of the missing high-frequency information for upscaling. Our Multi-input Multi-output Video Rescaling Network (MIMO-VRN) proposes a new strategy for downscaling and upscaling a group of video frames simultaneously. Not only do they outperform the image-based invertible model in terms of quantitative and qualitative results, but also show much improved upscaling quality than the video rescaling methods without joint optimization. To our best knowledge, this work is the first attempt at the joint optimization of video downscaling and upscaling.

Our work addresses long-term motion context issues for predicting future frames. To predict the future precisely, it is required to capture which long-term motion context (e.g., walking or running) the input motion (e.g., leg movement) belongs to. The bottlenecks arising when dealing with the long-term motion context are: (i) how to predict the long-term motion context naturally matching input sequences with limited dynamics, (ii) how to predict the long-term motion context with high-dimensionality (e.g., complex motion). To address the issues, we propose novel motion context-aware video prediction. To solve the bottleneck (i), we introduce a long-term motion context memory (LMC-Memory) with memory alignment learning. The proposed memory alignment learning enables to store long-term motion contexts into the memory and to match them with sequences including limited dynamics. As a result, the long-term context can be recalled from the limited input sequence. In addition, to resolve the bottleneck (ii), we propose memory query decomposition to store local motion context (i.e., low-dimensional dynamics) and recall the suitable local context for each local part of the input individually. It enables to boost the alignment effects of the memory. Experimental results show that the proposed method outperforms other sophisticated RNN-based methods, especially in long-term condition. Further, we validate the effectiveness of the proposed network designs by conducting ablation studies and memory feature analysis. The source code of this work is available.

In this paper, we propose a feature embedding based video object segmentation (VOS) method which is simple, fast and effective. The current VOS task involves two main challenges: object instance differentiation and cross-frame instance alignment. Most state-of-the-art matching based VOS methods simplify this task into a binary segmentation task and tackle each instance independently. In contrast, we decompose the VOS task into two subtasks: global embedding learning that segments foreground objects of each frame in a pixel-to-pixel manner, and instance feature embedding learning that separates instances. The outputs of these two subtasks are fused to obtain the final instance masks quickly and accurately. Through using the relation among different instances per-frame as well as temporal relation across different frames, the proposed network learns to differentiate multiple instances and associate them properly in one feed-forward manner. Extensive experimental results on the challenging DAVIS and Youtube-VOS datasets show that our method achieves better performances than most counterparts in each case.

Human pose estimation has achieved significant progress in recent years. However, most of the recent methods focus on improving accuracy using complicated models and ignoring real-time efficiency. To achieve a better trade-off between accuracy and efficiency, we propose a novel neural architecture search (NAS) method, termed ViPNAS, to search networks in both spatial and temporal levels for fast online video pose estimation. In the spatial level, we carefully design the search space with five different dimensions including network depth, width, kernel size, group number, and attentions. In the temporal level, we search from a series of temporal feature fusions to optimize the total accuracy and speed across multiple video frames. To the best of our knowledge, we are the first to search for the temporal feature fusion and automatic computation allocation in videos. Extensive experiments demonstrate the effectiveness of our approach on the challenging COCO2017 and PoseTrack2018 datasets. Our discovered model family, S-ViPNAS and T-ViPNAS, achieve significantly higher inference speed (CPU real-time) without sacrificing the accuracy compared to the previous state-of-the-art methods.

In this paper, we diagnose deep neural networks for 3D point cloud processing to explore utilities of different network architectures. We propose a number of hypotheses on the effects of specific network architectures on the representation capacity of DNNs. In order to prove the hypotheses, we design five metrics to diagnose various types of DNNs from the following perspectives, information discarding, information concentration, rotation robustness, adversarial robustness, and neighborhood inconsistency. We conduct comparative studies based on such metrics to verify the hypotheses. We further use the verified hypotheses to revise architectures of existing DNNs and improve their utilities. Experiments demonstrate the effectiveness of our method. The code will be released when this paper is accepted.

Self-supervised learning has gained prominence due to its efficacy at learning powerful representations from unlabelled data that achieve excellent performance on many challenging downstream tasks. However, supervision-free pre-text tasks are challenging to design and usually modality specific. Although there is a rich literature of self-supervised methods for either spatial (such as images) or temporal data (sound or text) modalities, a common pre-text task that benefits both modalities is largely missing. In this paper, we are interested in defining a self-supervised pre-text task for sketches and handwriting data. This data is uniquely characterised by its existence in dual modalities of rasterized images and vector coordinate sequences. We address and exploit this dual representation by proposing two novel cross-modal translation pre-text tasks for self-supervised feature learning: Vectorization and Rasterization. Vectorization learns to map image space to vector coordinates and rasterization maps vector coordinates to image space. We show that our learned encoder modules benefit both raster-based and vector-based downstream approaches to analysing hand-drawn data. Empirical evidence shows that our novel pre-text tasks surpass existing single and multi-modal self-supervision methods.

Accurately ranking the vast number of candidate detections is crucial for dense object detectors to achieve high performance. Prior work uses the classification score or a combination of classification and predicted localization scores to rank candidates. However, neither option results in a reliable ranking, thus degrading detection performance. In this paper, we propose to learn an Iou-Aware Classification Score (IACS) as a joint representation of object presence confidence and localization accuracy. We show that dense object detectors can achieve a more accurate ranking of candidate detections based on the IACS. We design a new loss function, named Varifocal Loss, to train a dense object detector to predict the IACS, and propose a new star-shaped bounding box feature representation for IACS prediction and bounding box refinement. Combining these two new components and a bounding box refinement branch, we build an IoU-aware dense object detector based on the FCOS+ATSS architecture, that we call VarifocalNet or VFNet for short. Extensive experiments on MS COCO show that our VFNet consistently surpasses the strong baseline by 2.0 AP with different backbones. Our best model VFNet-X-1200 with Res2Net-101-DCN achieves a single-model single-scale AP of 55.1 on COCO test-dev, which is state-of-the-art among various object detectors. Code is available at: https://github.com/hyz-xmaster/VarifocalNet.

Generative models able to synthesize layouts of different kinds (e.g. documents, user interfaces or furniture arrangements) are a useful tool to aid design processes and as a first step in the generation of synthetic data, among other tasks. We exploit the properties of self-attention layers to capture high level relationships between elements in a layout, and use these as the building blocks of the well-known Variational Autoencoder (VAE) formulation. Our proposed Variational Transformer Network (VTN) is capable of learning margins, alignments and other global design rules without explicit supervision. Layouts sampled from our model have a high degree of resemblance to the training data, while demonstrating appealing diversity. In an extensive evaluation on publicly available benchmarks for different layout types VTNs achieve state-of-the-art diversity and perceptual quality. Additionally, we show the capabilities of this method as part of a document layout detection pipeline.

Real-scanned point clouds are often incomplete due to viewpoint, occlusion, and noise. Existing point cloud completion methods tend to generate global shape skeletons and hence lack fine local details. Furthermore, they mostly learn a deterministic partial-to-complete mapping, but overlook structural relations in man-made objects. To tackle these challenges, this paper proposes a variational framework, Variational Relational point Completion network (VRCNet) with two appealing properties: 1) Probabilistic Modeling. In particular, we propose a dual-path architecture to enable principled probabilistic modeling across partial and complete clouds. One path consumes complete point clouds for reconstruction by learning a point VAE. The other path generates complete shapes for partial point clouds, whose embedded distribution is guided by distribution obtained from the reconstruction path during training. 2) Relational Enhancement. Specifically, we carefully design point self-attention kernel and point selective kernel module to exploit relational point features, which refines local shape details conditioned on the coarse completion. In addition, we contribute a multi-view partial point cloud dataset (MVP dataset) containing over 100,000 high-quality scans, which renders partial 3D shapes from 26 uniformly distributed camera poses for each 3D CAD model. Extensive experiments demonstrate that VRCNet outperforms state-of-the-art methods on all standard point cloud completion benchmarks. Notably, VRCNet shows great generalizability and robustness on real-world point cloud scans.