CVPR 2021 Papers

Training deep networks with limited labeled data while achieving a strong generalization ability is key in the quest to reduce human annotation efforts. This is the goal of semi-supervised learning, which exploits more widely available unlabeled data to complement small labeled data sets. In this paper, we propose a novel framework for discriminative pixel-level tasks using a generative model of both images and labels. Concretely, we learn a generative adversarial network that captures the joint image-label distribution and is trained efficiently using a large set of unlabeled images supplemented with only few labeled ones. We build our architecture on top of StyleGAN2, augmented with a label synthesis branch. Image labeling at test time is achieved by first embedding the target image into the joint latent space via an encoder network and test-time optimization, and then generating the label from the inferred embedding. We evaluate our approach in two important domains: medical image segmentation and part-based face segmentation. We demonstrate strong in-domain performance compared to several baselines, and are the first to showcase extreme out-of-domain generalization, such as transferring from CT to MRI in medical imaging, and photographs of real faces to paintings, sculptures, and even cartoons and animal faces.

Semantic Scene Completion aims at reconstructing a complete 3D scene with precise voxel-wise semantics from a single-view depth or RGBD image. It is a crucial but challenging problem for indoor scene understanding. In this work, we present a novel framework named Scene-Instance-Scene Network (SISNet), which takes advantages of both instance and scene level semantic information. Our method is capable of inferring fine-grained shape details as well as nearby objects whose semantic categories are easily mixedup. The key insight is that we decouple the instances from a coarsely completed semantic scene instead of a raw input image to guide the reconstruction of instances and the overall scene. SISNet conducts iterative scene-to-instance (SI) and instance-to-scene (IS) semantic completion. Specifically, the SI is able to encode objects' surrounding context for effectively decoupling instances from the scene and each instance could be voxelized into higher resolution to capture finer details. With IS, fine-grained instance information can be integrated back into the 3D scene and thus leads to more accurate semantic scene completion. Utilizing such an iterative mechanism, the scene and instance completion benefits each other to achieve higher completion accuracy. Extensively experiments show that our proposed method consistently outperforms state-of-the-art methods on both real NYU, NYUCAD and synthetic SUNCG-RGBD datasets. The code and the supplementary material will be available at https://github.com/yjcaimeow/SISNet.

Few-shot object detection is an imperative and long-lasting problem due to the inherent long-tail distribution of real-world data. Its performance is largely affected by the data scarcity of novel classes. But the semantic relation between the novel classes and the base classes is constant regardless of the data availability. In this work, we investigate utilizing this semantic relation together with the visual information and introduce explicit relation reasoning into the learning of novel object detection. Specifically, we represent each class concept by a semantic embedding learned from a large corpus of text. The detector is trained to project the image representations of objects into this embedding space. We also identify the problems of trivially using the raw embeddings with a heuristic knowledge graph and propose to augment the embeddings with a dynamic relation graph. As a result, our few-shot detector, termed SRR-FSD, is robust and stable to the variation of shots of novel objects. Experiments show that SRR-FSD can achieve competitive results at higher shots, and more importantly, a significantly better performance given both lower explicit and implicit shots. The benchmark protocol with implicit shots removed from the pretrained classification dataset can serve as a more realistic setting for future research.

Despite the recent progress of generative adversarial networks (GANs) at synthesizing photo-realistic images, producing complex urban scenes remains a challenging problem. Previous works break down scene generation into two consecutive phases: unconditional semantic layout synthesis and image synthesis conditioned on layouts. In this work, we propose to condition layout generation as well for higher semantic control: given a vector of class proportions, we generate layouts with matching composition. To this end, we introduce a conditional framework with novel architecture designs and learning objectives, which effectively accommodates class proportions to guide the scene generation process. The proposed architecture also allows partial layout editing with interesting applications. Thanks to the semantic control, we can produce layouts close to the real distribution, helping enhance the whole scene generation process. On different metrics and urban scene benchmarks, our models outperform existing baselines. Moreover, we demonstrate the merit of our approach for data augmentation: semantic segmenters trained on real layout-image pairs along with additional ones generated by our approach outperform models only trained on real pairs.

Natural image matting separates the foreground from background in fractional occupancy which can be caused by highly transparent objects, complex foreground (e.g., net or tree), and/or objects containing very fine details (e.g., hairs). Although conventional matting formulation can be applied to all of the above cases, no previous work has attempted to reason the underlying causes of matting due to various foreground semantics. We show how to obtain better alpha mattes by incorporating into our framework semantic classification of matting regions. Specifically, we consider and learn 20 classes of matting patterns, and propose to extend the conventional trimap to semantic trimap. The proposed semantic trimap can be obtained automatically through patch structure analysis within trimap regions. Meanwhile, we learn a multi-class discriminator to regularize the alpha prediction at semantic level, and content-sensitive weights to balance different regularization losses. Experiments on multiple benchmarks show that our method outperforms other methods and has achieved the most competitive state-of-the-art performance. Finally, we contribute a large-scale Semantic Image Matting Dataset with careful consideration of data balancing across different semantic classes. Code and dataset will be released.

Recent work on audio-visual navigation assumes a constantly-sounding target and restricts the role of audio to signaling the target's position. We introduce semantic audio-visual navigation, where objects in the environment make sounds consistent with their semantic meaning (e.g., toilet flushing, door creaking) and acoustic events are sporadic or short in duration. We propose a transformer-based model to tackle this new semantic AudioGoal task, incorporating an inferred goal descriptor that captures both spatial and semantic properties of the target. Our model's persistent multimodal memory enables it to reach the goal even long after the acoustic event stops. In support of the new task, we also expand the SoundSpaces audio simulations to provide semantically grounded sounds for an array of objects in Matterport3D. Our method strongly outperforms existing audio-visual navigation methods by learning to associate semantic, acoustic, and visual cues. Project page: http://vision.cs.utexas.edu/projects/semantic-audio-visual-navigation.

Graph convolution networks (GCNs) are a powerful deep learning approach and have been successfully applied to representation learning on graphs in a variety of real-world applications. Despite their success, two fundamental weaknesses of GCNs limit their ability to represent graph-structured data: poor performance when labeled data are severely scarce and indistinguishable features when more layers are stacked. In this paper, we propose a simple yet effective Self-Supervised Semantic Alignment Graph Convolution Network (SelfSAGCN), which consists of two crux techniques: Identity Aggregation and Semantic Alignment, to overcome these weaknesses. The behind basic idea is the node features in the same class but learned from semantic and graph structural aspects respectively, are expected to be mapped nearby. Specifically, the Identity Aggregation is applied to extract semantic features from labeled nodes, the Semantic Alignment is utilized to align node features obtained from different aspects using the class central similarity. In this way, the over-smoothing phenomenon is alleviated, while the similarities between the unlabeled features and labeled ones from the same class are enhanced. Experimental results on five popular datasets show that the proposed SelfSAGCN outperforms state-of-the-art methods on various classification tasks.

We propose SelfDoc, a task-agnostic pre-training framework for document image understanding. Because documents are multimodal and are intended for sequential reading, our framework exploits the positional, textual, and visual information of every semantically meaningful component in a document, and it models the contextualization between each block of content. Unlike existing document pre-training models, our model is coarse-grained instead of treating individual words as input, therefore avoiding an overly fine-grained with excessive contextualization. Beyond that, we introduce cross-modal learning in the model pre-training phase to fully leverage multimodal information from unlabeled documents. For downstream usage, we propose a novel modality-adaptive attention mechanism for multimodal feature fusion by adaptively emphasizing language and vision signals. Our framework benefits from self-supervised pre-training on documents without requiring annotations by a feature masking training strategy. It achieves superior performance on multiple downstream tasks with significantly fewer document images used in the pre-training stage compared to previous works.

A common practice in unsupervised representation learning is to use labeled data to evaluate the quality of the learned representations. This supervised evaluation is then used to guide critical aspects of the training process such as selecting the data augmentation policy. However, guiding an unsupervised training process through supervised evaluations is not possible for real-world data that does not actually contain labels (which may be the case, for example, in privacy sensitive fields such as medical imaging). Therefore, in this work we show that evaluating the learned representations with a self-supervised image rotation task is highly correlated with a standard set of supervised evaluations (rank correlation > 0.94). We establish this correlation across hundreds of augmentation policies, training settings, and network architectures and provide an algorithm (SelfAugment) to automatically and efficiently select augmentation policies without using supervised evaluations. Despite not using any labeled data, the learned augmentation policies perform comparably with augmentation policies that were determined using exhaustive supervised evaluations.

The goal is to use Wasserstein metric to provide pseudo labels for the unlabeled images to train a Convolutional Neural Networks (CNN) in a Semi-Supervised Learning (SSL) manner for the classification task. The basic premise in our method is that the discrepancy between two discrete empirical measures (e.g., clusters) which come from the same or similar distribution is expected to be less than the case where these measures come from completely two different distributions. In our proposed method, we first pre-train our CNN using a self-supervised learning method to make a cluster assumption on the unlabeled images. Next, inspired by the Wasserstein metric which considers the geometry of the metric space to provide a natural notion of similarity between discrete empirical measures, we leverage it to cluster the unlabeled images and then match the clusters to their similar class of labeled images to provide a pseudo label for the data within each cluster. We have evaluated and compared our method with state-of-the-art SSL methods on the standard datasets to demonstrate its effectiveness.

We address the problem of novel view synthesis (NVS) from a few sparse source view images. Conventional image-based rendering methods estimate scene geometry and synthesize novel views in two separate steps. However, erroneous geometry estimation will decrease NVS performance as view synthesis highly depends on the quality of estimated scene geometry. In this paper, we propose an end-to-end NVS framework to eliminate the error propagation issue. To be specific, we construct a volume under the target view and design a source-view visibility estimation (SVE) module to determine the visibility of the target-view voxels in each source view. Next, we aggregate the visibility of all source views to achieve a consensus volume. Each voxel in the consensus volume indicates a surface existence probability. Then, we present a soft ray-casting (SRC) mechanism to find the most front surface in the target view (i.e. depth). Specifically, our SRC traverses the consensus volume along viewing rays and then estimates a depth probability distribution. We then warp and aggregate source view pixels to synthesize a novel view based on the estimated source-view visibility and target-view depth. At last, our network is trained in an end-to-end self-supervised fashion, thus significantly alleviating error accumulation in view synthesis. Experimental results demonstrate that our method generates novel views in higher quality compared to the state-of-the-art.

A key challenge in self-supervised video representation learning is how to effectively capture motion information besides context bias. While most existing works implicitly achieve this with video-specific pretext tasks (e.g., predicting clip orders, time arrows, and paces), we develop a method that explicitly decouples motion supervision from context bias through a carefully designed pretext task. Specifically, we take the key frames and motion vectors in compressed videos (e.g., in H.264 format) as the supervision sources for context and motion, respectively, which can be efficiently extracted at over 500 fps on CPU. Then we design two pretext tasks that are jointly optimized: a context matching task where a pairwise contrastive loss is cast between video clip and key frame features; and a motion prediction task where clip features, passed through an encoder-decoder network, are used to estimate motion features in a near future. These two tasks use a shared video backbone and separate MLP heads. Experiments show that our approach improves the quality of the learned video representation over previous works, where we obtain absolute gains of 16.0% and 11.1% in video retrieval recall on UCF101 and HMDB51, respectively. Moreover, we find the motion prediction to be a strong regularization for video networks, where using it as an auxiliary task improves the accuracy of action recognition with a margin of 7.4% 13.8%.

Most existing unsupervised video hashing methods are built on unidirectional models with less reliable training objectives, which underuse the correlations among frames and the similarity structure between videos. To enable efficient scalable video retrieval, we propose a self-supervised video Hashing method based on Bidirectional Transformers (BTH). Based on the encoder-decoder structure of transformers, we design a visual cloze task to fully exploit the bidirectional correlations between frames. To unveil the similarity structure between unlabeled video data, we further develop a similarity reconstruction task by establishing reliable and effective similarity connections in the video space. Furthermore, we develop a cluster assignment task to exploit the structural statistics of the whole dataset such that more discriminative binary codes can be learned. Extensive experiments implemented on three public benchmark datasets, FCVID, ActivityNet and YFCC, demonstrate the superiority of our proposed approach.

A video can be represented by the composition of appearance and motion. Appearance (or content) expresses the information invariant throughout time, and motion describes the time-variant movement. Here, we propose self-supervised approaches for video Generative Adversarial Networks (GANs) to achieve the appearance consistency and motion coherency in videos. Specifically, the dual discriminators for image and video individually learn to solve their own pretext tasks; appearance contrastive learning and temporal structure puzzle. The proposed tasks enable the discriminators to learn representations of appearance and temporal context, and force the generator to synthesize videos with consistent appearance and natural flow of motions. Extensive experiments in facial expression and human action public benchmarks show that our method outperforms the state-of-the-art video GANs. Moreover, consistent improvements regardless of the architecture of video GANs confirm that our framework is generic.

In this paper we propose a system consisting of a modular network and a trajectory planner. The network simultaneously predicts Occupancy Grid Maps (OGMs) and estimates space-time cost maps (CMs) corresponding to the areas around the vehicle. The trajectory planner computes the cost of a set of predefined trajectories and chooses the one with the lowest cost. Training this network is done in a self-supervised manner which desirably do not require any labeled data. The proposed training objective takes into account the accuracy of OGM predictions as well as contextual information and human driver behavior. Training these modules end-to-end makes each module aware of the errors caused by the other components of the system. We show that our proposed method can lead to the selection of low cost trajectories with a low collision rate and road violation in fairly long planning horizons.

Autonomous driving can benefit from motion behavior comprehension when interacting with diverse traffic participants in highly dynamic environments. Recently, there has been a growing interest in estimating class-agnostic motion directly from point clouds. Current motion estimation methods usually require vast amount of annotated training data from self-driving scenes. However, manually labeling point clouds is notoriously difficult, error-prone and time-consuming. In this paper, we seek to answer the research question of whether the abundant unlabeled data collections can be utilized for accurate and efficient motion learning. To this end, we propose a learning framework that leverages free supervisory signals from point clouds and paired camera images to estimate motion purely via self-supervision. Our model involves a point cloud based structural consistency augmented with probabilistic motion masking as well as a cross-sensor motion regularization to realize the desired self-supervision. Experiments reveal that our approach performs competitively to supervised methods, and achieves the state-of-the-art result when combining our self-supervised model with supervised fine-tuning.

Estimating 3D scene flow from a sequence of monocular images has been gaining increased attention due to the simple, economical capture setup. Owing to the severe ill-posedness of the problem, the accuracy of current methods has been limited, especially that of efficient, real-time approaches. In this paper, we introduce a multi-frame monocular scene flow network based on self-supervised learning, improving the accuracy over previous networks while retaining real-time efficiency. Based on an advanced two-frame baseline with a split-decoder design, we propose (i) a multi-frame model using a triple frame input and convolutional LSTM connections, (ii) an occlusion-aware census loss for better accuracy, and (iii) a gradient detaching strategy to improve training stability. On the KITTI dataset, we observe state-of-the-art accuracy among monocular scene flow methods based on self-supervised learning.

Motions are reflected in videos as the movement of pixels, and actions are essentially patterns of inconsistent motions between the foreground and the background. To well distinguish the actions, especially those with complicated spatio-temporal interactions, correctly locating the prominent motion areas is of crucial importance. However, most motion information in existing videos are difficult to label and training a model with good motion representations with supervision will thus require a large amount of human labour for annotation. In this paper, we address this problem by self-supervised learning. Specifically, we propose to learn Motion from Static Images (MoSI). The model learns to encode motion information by classifying pseudo motions generated by MoSI. We furthermore introduce a static mask in pseudo motions to create local motion patterns, which forces the model to additionally locate notable motion areas for the correct classification.We demonstrate that MoSI can discover regions with large motion even without fine-tuning on the downstream datasets. As a result, the learned motion representations boost the performance of tasks requiring understanding of complex scenes and motions, i.e., action recognition. Extensive experiments show the consistent and transferable improvements achieved by MoSI. Codes will be soon released.

While recent pre-training tasks on 2D images have proven very successful for transfer learning, pre-training for 3D data remains challenging. In this work, we introduce a general method for 3D self-supervised representation learning that 1) remains agnostic to the underlying neural network architecture, and 2) specifically leverages the geometric nature of 3D point cloud data. The proposed task softly segments 3D points into a discrete number of geometric partitions. A self-supervised loss is formed under the interpretation that these soft partitions implicitly parameterize a latent Gaussian Mixture Model (GMM), and that this generative model establishes a data likelihood function. Our pretext task can therefore be viewed in terms of an encoder-decoder paradigm that squeezes learned representations through an implicitly defined parametric discrete generative model bottleneck. We show that any existing neural network architecture designed for supervised point cloud segmentation can be repurposed for the proposed unsupervised pretext task. By maximizing data likelihood with respect to the soft partitions formed by the unsupervised point-wise segmentation network, learned representations are encouraged to contain compositionally rich geometric information. In tests, we show that our method naturally induces semantic separation in feature space, resulting in state-of-the-art performance on downstream applications like model classification and semantic segmentation.