CVPR 2022 Papers

Standard visual localization methods build a priori 3D model of a scene which is used to establish correspondences against the 2D keypoints in a query image. Storing these pre-built 3D scene models can be prohibitively expensive for large-scale environments, especially on mobile devices with limited storage and communication bandwidth. We design a novel framework that compresses a scene while still maintaining localization accuracy. The scene is compressed in three stages: first, the database frames are clustered using pairwise co-visibility information. Then, a learned point selection module prunes the points in each cluster taking into account the final pose estimation accuracy. In the final stage, the features of the selected points are further compressed using learned quantization. Query image registration is done using only the compressed scene points. To the best of our knowledge, we are the first to propose learned scene compression for visual localization. We also demonstrate the effectiveness and efficiency of our method on various outdoor datasets where it can perform accurate localization with low memory consumption.

A classical problem in computer vision is to infer a 3D scene representation from few images that can be used to render novel views at interactive rates. Previous work focuses on reconstructing pre-defined 3D representations, e.g. textured meshes, or implicit representations, e.g. radiance fields, and often requires input images with precise camera poses and long processing times for each novel scene. In this work, we propose the Scene Representation Transformer (SRT), a method which processes posed or unposed RGB images of a new area, infers a "set-latent scene representation", and synthesises novel views, all in a single feed-forward pass. To calculate the scene representation, we propose a generalization of the Vision Transformer to sets of images, enabling global information integration, and hence 3D reasoning. An efficient decoder transformer parameterizes the light field by attending into the scene representation to render novel views. Learning is supervised end-to-end by minimizing a novel-view reconstruction error. We show that this method outperforms recent baselines in terms of PSNR and speed on synthetic datasets, including a new dataset created for the paper. Further, we demonstrate that SRT scales to support interactive visualization and semantic segmentation of real-world outdoor environments using Street View imagery.

In this paper, we address the task of semantics-guided image outpainting, which is to complete an image by generating semantically practical content. Different from most existing image outpainting works, we approach the above task by understanding and completing image semantics at the scene graph level. In particular, we propose a novel network of Scene Graph Transformer (SGT), which is designed to take node and edge features as inputs for modeling the associated structural information. To better understand and process graph-based inputs, our SGT uniquely performs feature attention at both node and edge levels. While the former views edges as relationship regularization, the latter observes the co-occurrence of nodes for guiding the attention process. We demonstrate that, given a partial input image with its layout and scene graph, our SGT can be applied for scene graph expansion and its conversion to a complete layout. Following state-of-the-art layout-to-image conversions works, the task of image outpainting can be completed with sufficient and practical semantics introduced. Extensive experiments are conducted on the datasets of MS-COCO and Visual Genome, which quantitatively and qualitatively confirm the effectiveness of our proposed SGT and outpainting frameworks.

A long-term video, such as a movie or TV show, is composed of various scenes, each of which represents a series of shots sharing the same semantic story. Spotting the correct scene boundary from the long-term video is a challenging task, since a model must understand the storyline of the video to figure out where a scene starts and ends. To this end, we propose an effective Self-Supervised Learning (SSL) framework to learn better shot representations from unlabeled long-term videos. More specifically, we present an SSL scheme to achieve scene consistency, while exploring considerable data augmentation and shuffling methods to boost the model generalizability. Instead of explicitly learning the scene boundary features as in the previous methods, we introduce a vanilla temporal model with less inductive bias to verify the quality of the shot features. Our method achieves the state-of-the-art performance on the task of Video Scene Segmentation. Additionally, we suggest a more fair and reasonable benchmark to evaluate the performance of Video Scene Segmentation methods. The code is made available.

Trajectory prediction is a critical functionality of autonomous systems that share environments with uncontrolled agents, one prominent example being self-driving vehicles. Currently, most prediction methods do not enforce scene consistency, i.e., there are a substantial amount of self-collisions between predicted trajectories of different agents in the scene. Moreover, many approaches generate individual trajectory predictions per agent instead of joint trajectory predictions of the whole scene, which makes downstream planning difficult. In this work, we present ScePT, a policy planning-based trajectory prediction model that generates accurate, scene-consistent trajectory predictions suitable for autonomous system motion planning. It explicitly enforces scene consistency and learns an agent interaction policy that can be used for conditional prediction. Experiments on multiple real-world pedestrians and autonomous vehicle datasets show that ScePT matches current state-of-the-art prediction accuracy with significantly improved scene consistency. We also demonstrate ScePT's ability to work with a downstream contingency planner.

Cameras on portable devices are manufactured with a rolling-shutter (RS) mechanism, where the image rows (aka. scanlines) are read out sequentially. The unknown camera motions during the imaging process cause the so-called RS effects which are solved by motion assumptions in the literature. In this work, we give a solution to the absolute pose problem free of motion assumptions. We categorically demonstrate that the only requirement is motion smoothness instead of stronger constraints on the camera motion. To this end, we propose a novel mathematical abstraction for RS cameras observing a planar scene, called the scanline-homography, a 3x2 matrix with 5 DOFs. We establish the relationship between a scanline-homography and the corresponding plane-homography, a 3x3 matrix with 6 DOFs assuming the camera is calibrated. We estimate the scanline-homographies of an RS frame using a smooth image warp powered by B-Splines, and recover the plane-homographies afterwards to obtain the scanline-poses based on motion smoothness. We back our claims with various experiments. Code and new datasets: https://bitbucket.org/clermontferrand/planarscanlinehomography/src/master/.

We propose a new 3D spatial understanding task of 3D Question Answering (3D-QA). In the 3D-QA task, models receive visual information from the entire 3D scene of the rich RGB-D indoor scan and answer the given textual questions about the 3D scene. Unlike the 2D-question answering of VQA, the conventional 2D-QA models suffer from problems with spatial understanding of object alignment and directions and fail the object identification from the textual questions in 3D-QA. We propose a baseline model for 3D-QA, named ScanQA model, where the model learns a fused descriptor from 3D object proposals and encoded sentence embeddings. This learned descriptor correlates the language expressions with the underlying geometric features of the 3D scan and facilitates the regression of 3D bounding boxes to determine described objects in textual questions. We collected human-edited question-answer pairs with free-form answers that are grounded to 3D objects in each 3D scene. Our new ScanQA dataset contains over 40K question-answer pairs from the 800 indoor scenes drawn from the ScanNet dataset. To the best of our knowledge, ScanQA is the first large-scale effort to perform object-grounded question-answering in 3D environments.

Vision Transformers (ViTs) and their multi-scale and hierarchical variations have been successful at capturing image representations but their use has been generally studied for low-resolution images (e.g. - 256x256, 384x384). For gigapixel whole-slide imaging (WSI) in computational pathology, WSIs can be as large as 150000x150000 pixels at 20x magnification and exhibit a hierarchical structure of visual tokens across varying resolutions: from 16x16 images capture spatial patterns among cells, to 4096x4096 images characterizing interactions within the tissue microenvironment. We introduce a new ViT architecture called the Hierarchical Image Pyramid Transformer (HIPT), which leverages the natural hierarchical structure inherent in WSIs using two levels of self-supervised learning to learn high-resolution image representations. HIPT is pretrained across 33 cancer types using 10,678 gigapixel WSIs, 408,218 4096x4096 images, and 104M 256x256 images. We benchmark HIPT representations on 9 slide-level tasks, and demonstrate that: 1) HIPT with hierarchical pretraining outperforms current state-of-the-art methods for cancer subtyping and survival prediction, 2) self-supervised ViTs are able to model important inductive biases about the hierarchical structure of phenotypes in the tumor microenvironment.

Attention-based neural networks such as the Vision Transformer (ViT) have recently attained state-of-the-art results on many computer vision benchmarks. Scale is a primary ingredient in attaining excellent results, therefore, understanding a model's scaling properties is a key to designing future generations effectively. While the laws for scaling Transformer language models have been studied, it is unknown how Vision Transformers scale. To address this, we scale ViT models and data, both up and down, and characterize the relationships between error rate, data, and compute. Along the way, we refine the architecture and training of ViT, reducing memory consumption and increasing accuracy of the resulting models. As a result, we successfully train a ViT model with two billion parameters, which attains a new state-of-the-art on ImageNet of 90.45% top-1 accuracy. The model also performs well for few-shot transfer, for example, reaching 84.86% top-1 accuracy on ImageNet with only 10 examples per class.

We revisit large kernel design in modern convolutional neural networks (CNNs). Inspired by recent advances in vision transformers (ViTs), in this paper, we demonstrate that using a few large convolutional kernels instead of a stack of small kernels could be a more powerful paradigm. We suggested five guidelines, e.g., applying re-parameterized large depth-wise convolutions, to design efficient high-performance large-kernel CNNs. Following the guidelines, we propose RepLKNet, a pure CNN architecture whose kernel size is as large as 31x31, in contrast to commonly used 3x3. RepLKNet greatly closes the performance gap between CNNs and ViTs, e.g., achieving comparable or superior results than Swin Transformer on ImageNet and a few typical downstream tasks, with lower latency. RepLKNet also shows nice scalability to big data and large models, obtaining 87.8% top-1 accuracy on ImageNet and 56.0% mIoU on ADE20K, which is very competitive among the state-of-the-arts with similar model sizes. Our study further reveals that, in contrast to small-kernel CNNs, large-kernel CNNs have much larger effective receptive fields and higher shape bias rather than texture bias. Code & models at https://github.com/megvii-research/RepLKNet.

In recent years, we have witnessed significant performance boost in the image captioning task based on vision-language pre-training (VLP). Scale is believed to be an important factor for this advance. However, most existing work only focuses on pre-training transformers with moderate sizes (e.g., 12 or 24 layers) on roughly 4 million images. In this paper, we present LEMON, a LargE-scale iMage captiONer, and provide the first empirical study on the scaling behavior of VLP for image captioning. We use the state-of-the-art VinVL model as our reference model, which consists of an image feature extractor and a transformer model, and scale the transformer both up and down, with model sizes ranging from 13 to 675 million parameters. In terms of data, we conduct experiments with up to 200 million image-text pairs which are automatically collected from web based on the alt attribute of the image (dubbed as ALT200M). Extensive analysis helps to characterize the performance trend as the model size and the pre-training data size increase. We also compare different training recipes, especially for training on large-scale noisy data. As a result, LEMON achieves new state of the arts on several major image captioning benchmarks, including COCO Caption, nocaps, and Conceptual Captions. We also show LEMON can generate captions with long-tail visual concepts when used in a zero-shot manner.

In this paper, we address the problem of estimating scale factors between images. We formulate the scale estimation problem as a prediction of a probability distribution over scale factors. We design a new architecture, ScaleNet, that exploits dilated convolutions as well as self- and cross-correlation layers to predict the scale between images. We demonstrate that rectifying images with estimated scales leads to significant performance improvements for various tasks and methods. Specifically, we show how ScaleNet can be combined with sparse local features and dense correspondence networks to improve camera pose estimation, 3D reconstruction, or dense geometric matching in different benchmarks and datasets. We provide an extensive evaluation on several tasks, and analyze the computational overhead of ScaleNet. The code, evaluation protocols, and trained models are publicly available.

Recent Semi-Supervised Object Detection (SS-OD) methods are mainly based on self-training, i.e., generating hard pseudo-labels by a teacher model on unlabeled data as supervisory signals. Although they achieved certain success, the limited labeled data in semi-supervised learning scales up the challenges of object detection. We analyze the challenges these methods meet with the empirical experiment results. We find that the massive False Negative samples and inferior localization precision lack consideration. Besides, the large variance of object sizes and class imbalance (i.e., the extreme ratio between background and object) hinder the performance of prior arts. Further, we overcome these challenges by introducing a novel approach, Scale-Equivalent Distillation (SED), which is a simple yet effective end-to-end knowledge distillation framework robust to large object size variance and class imbalance. SED has several appealing benefits compared to the previous works. (1) SED imposes a consistency regularization to handle the large scale variance problem. (2) SED alleviates the noise problem from the False Negative samples and inferior localization precision. (3) A re-weighting strategy can implicitly screen the potential foreground regions of the unlabeled data to reduce the effect of class imbalance. Extensive experiments show that SED consistently outperforms the recent state-of-the-art methods on different datasets with significant margins. For example, it surpasses the supervised counterpart by more than 10 mAP when using 5% and 10% labeled data on MS-COCO.

Noisy training set usually leads to the degradation of generalization and robustness of neural networks. In this paper, we propose using a theoretically guaranteed noisy label detection framework to detect and remove noisy data for Learning with Noisy Labels (LNL). Specifically, we design a penalized regression to model the linear relation between network features and one-hot labels, where the noisy data are identified by the non-zero mean shift parameters solved in the regression model. To make the framework scalable to datasets that contain a large number of categories and training data, we propose a split algorithm to divide the whole training set into small pieces that can be solved by the penalized regression in parallel, leading to the Scalable Penalized Regression (SPR) framework. We provide the non-asymptotic probabilistic condition for SPR to correctly identify the noisy data. While SPR can be regarded as a sample selection module for standard supervised training pipeline, we further combine it with semi-supervised algorithm to further exploit the support of noisy data as unlabeled data. Experimental results on several benchmark datasets and real-world noisy datasets show the effectiveness of our framework. Our code and pretrained models are released at https://github.com/Yikai-Wang/SPR-LNL.

Weakly supervised object detection (WSOD) has recently attracted much attention. However, the lack of bounding-box supervision makes its accuracy much lower than fully supervised object detection (FSOD), and currently modern FSOD techniques cannot be applied to WSOD. To bridge the performance and technical gaps between WSOD and FSOD, this paper proposes a new framework, Salvage of Supervision (SoS), with the key idea being to harness every potentially useful supervisory signal in WSOD: the weak image-level labels, the pseudo-labels, and the power of semi-supervised object detection. This paper shows that each type of supervisory signal brings in notable improvements, outperforms existing WSOD methods (which mainly use only the weak labels) by large margins. The proposed SoS-WSOD method also have the ability to freely use modern FSOD techniques. SoS-WSOD achieves 64.4 mAP50 on VOC2007, 61.9 mAP50 on VOC2012 and 16.6 mAP50:95 on MS-COCO, and also has fast inference speed. Ablations and visualization further verify the effectiveness of SoS.

Due to the limited utilization of temporal relations in video re-id, the frame-level attention regions of mainstream methods are partial and highly similar. To address this problem, we propose a Salient-to-Broad Module (SBM) to enlarge the attention regions gradually. Specifically, in SBM, while the previous frames have focused on the most salient regions, the later frames tend to focus on broader regions. In this way, the additional information in broad regions can supplement salient regions, incurring more powerful video-level representations. To further improve SBM, an Integration-and-Distribution Module (IDM) is introduced to enhance frame-level representations. IDM first integrates features from the entire feature space and then distributes the integrated features to each spatial location. SBM and IDM are mutually beneficial since they enhance the representations from video-level and framelevel, respectively. Extensive experiments on four prevalent benchmarks demonstrate the effectiveness and superiority of our method. The source code is available at https://github.com/baist/SINet.

Deep semi-supervised learning (SSL) methods aim to take advantage of abundant unlabeled data to improve the algorithm performance. In this paper, we consider the problem of safe SSL scenario where unseen-class instances appear in the unlabeled data. This setting is essential and commonly appears in a variety of real applications. One intuitive solution is removing these unseen-class instances after detecting them during the SSL process. Nevertheless, the performance of unseen-class identification is limited by the small number of labeled data and ignoring the availability of unlabeled data. To take advantage of these unseen-class data and ensure performance, we propose a safe SSL method called SAFE-STUDENT from the teacher-student view. Firstly, a new scoring function called energy-discrepancy (ED) is proposed to help the teacher model improve the security of instances selection. Then, a novel unseen-class label distribution learning mechanism mitigates the unseen-class perturbation by calibrating the unseen-class label distribution. Finally, we propose an iterative optimization strategy to facilitate teacher-student network learning. Extensive studies on several representative datasets show that SAFE-STUDENT remarkably outperforms the state-of-the-art, verifying the feasibility and robustness of our method in the under-explored problem.

Unsupervised Domain Adaptation (UDA) aims to leverage a label-rich source domain to solve tasks on a related unlabeled target domain. It is a challenging problem especially when a large domain gap lies between the source and target domains. In this paper we propose a novel solution named SSRT (Safe Self-Refinement for Transformer-based domain adaptation), which brings improvement from two aspects. First, encouraged by the success of vision transformers in various vision tasks, we arm SSRT with a transformer backbone. We find that the combination of vision transformer with simple adversarial adaptation surpasses best reported Convolutional Neural Network (CNN)-based results on the challenging DomainNet benchmark, showing its strong transferable feature representation. Second, to reduce the risk of model collapse and improve the effectiveness of knowledge transfer between domains with large gaps, we propose a Safe Self-Refinement strategy. Specifically, SSRT utilizes predictions of perturbed target domain data to refine the model. Since the model capacity of vision transformer is large and predictions in such challenging tasks can be noisy, a safe training mechanism is designed to adaptively adjust learning configuration. Extensive evaluations are conducted on several widely tested UDA benchmarks and SSRT achieves consistently the best performances, including 85.43% on Office-Home, 88.76% on VisDA-2017 and 45.2% on DomainNet.