CVPR 2021 Papers

Given an incomplete image without additional constraint, image inpainting natively allows for multiple solutions as long as they appear plausible. Recently, multiple-solution inpainting methods have been proposed and shown the potential of generating diverse results. However, these methods have difficulty in ensuring the quality of each solution, e.g. they produce distorted structure and/or blurry texture. We propose a two-stage model for diverse inpainting, where the first stage generates multiple coarse results each of which has a different structure, and the second stage refines each coarse result separately by augmenting texture. The proposed model is inspired by the hierarchical vector quantized variational auto-encoder (VQ-VAE), whose hierarchical architecture disentangles structural and textural information. In addition, the vector quantization in VQ-VAE enables autoregressive modeling of the discrete distribution over the structural information. Sampling from the distribution can easily generate diverse and high-quality structures, making up the first stage of our model. In the second stage, we propose a structural attention module inside the texture generation network, where the module utilizes the structural information to capture distant correlations. We further reuse the VQ-VAE to calculate two feature losses, which help improve structure coherence and texture realism, respectively. Experimental results on CelebA-HQ, Places2, and ImageNet datasets show that our method not only enhances the diversity of the inpainting solutions but also improves the visual quality of the generated multiple images. Code and models are available at: https://github.com/USTC-JialunPeng/Diverse-Structure-Inpainting.

Despite learning-based visual odometry (VO) has shown impressive results in recent years, the pretrained networks may easily collapse in unseen environments. The large domain gap between training and testing data makes them difficult to generalize to new scenes. In this paper, we propose an online adaptation framework for deep VO with the assistance of scene-agnostic geometric computations and Bayesian inference. In contrast to learning-based pose estimation, our method solves pose from optical flow and depth while the single-view depth estimation is continuously improved with new observations by online learned uncertainties. Meanwhile, an online learned photometric uncertainty is used for further depth and pose optimization by a differentiable Gauss-Newton layer. Our method enables fast adaptation of deep VO networks to unseen environments in a self-supervised manner. Extensive experiments including Cityscapes to KITTI and outdoor KITTI to indoor TUM demonstrate that our method achieves state-of-the-art generalization ability among self-supervised VO methods.

Current face forgery detection methods achieve high accuracy under the within-database scenario where training and testing forgeries are synthesized by the same algorithm. However, few of them gain satisfying performance under the cross-database scenario where training and testing forgeries are synthesized by different algorithms. In this paper, we find that current CNN-based detectors tend to overfit to method-specific color textures and thus fail to generalize. Observing that image noises remove color textures and expose discrepancies between authentic and tampered regions, we propose to utilize the high-frequency noises for face forgery detection. We carefully devise three functional modules to take full advantage of the high-frequency features. The first is the multi-scale high-frequency feature extraction module that extracts high-frequency noises at multiple scales and composes a novel modality. The second is the residual-guided spatial attention module that guides the low-level RGB feature extractor to concentrate more on forgery traces from a new perspective. The last is the cross-modality attention module that leverages the correlation between the two complementary modalities to promote feature learning for each other. Comprehensive evaluations on several benchmark databases corroborate the superior generalization performance of our proposed method.

Localization Quality Estimation (LQE) is crucial and popular in the recent advancement of dense object detectors since it can provide accurate ranking scores that benefit the Non-Maximum Suppression processing and improve detection performance. As a common practice, most existing methods predict LQE scores through vanilla convolutional features shared with object classification or bounding box regression. In this paper, we explore a completely novel and different perspective to perform LQE -- based on the learned distributions of the four parameters of the bounding box. The bounding box distributions are inspired and introduced as "General Distribution" in GFLV1, which describes the uncertainty of the predicted bounding boxes well. Such a property makes the distribution statistics of a bounding box highly correlated to its real localization quality. Specifically, a bounding box distribution with a sharp peak usually corresponds to high localization quality, and vice versa. By leveraging the close correlation between distribution statistics and the real localization quality, we develop a considerably lightweight Distribution-Guided Quality Predictor (DGQP) for reliable LQE based on GFLV1, thus producing GFLV2. To our best knowledge, it is the first attempt in object detection to use a highly relevant, statistical representation to facilitate LQE. Extensive experiments demonstrate the effectiveness of our method. Notably, GFLV2 (ResNet-101) achieves 46.2 AP at 14.6 FPS, surpassing the previous state-of-the-art ATSS baseline (43.6 AP at 14.6 FPS) by absolute 2.6 AP on COCO \tt test-dev , without sacrificing the efficiency both in training and inference.

Learning object detection from few examples recently emerged to deal with data-limited situations. While most previous works merely focus on the performance on few-shot categories, we claim that the ability to detect all classes is crucial as test samples may contain any instances in realistic applications, which requires the few-shot detector to learn new concepts without forgetting. Through analysis on transfer learning based methods, some neglected but beneficial properties are utilized to design a simple yet effective few-shot detector, Retentive R-CNN. It consists of Bias-Balanced RPN to debias the pretrained RPN and Re-detector to find few-shot class objects without forgetting previous knowledge. Extensive experiments on few-shot detection benchmarks show that Retentive R-CNN significantly outperforms state-of-the-art methods on overall performance among all settings as it can achieve competitive results on few-shot classes and does not degrade on base class performance at all. Our approach has demonstrated that the long desired never-forgetting learner is available in object detection.

Many variants of unsupervised domain adaptation (UDA) problems have been proposed and solved individually. Its side effect is that a method that works for one variant is often ineffective for or not even applicable to another, which has prevented practical applications. In this paper, we give a general representation of UDA problems, named Generalized Domain Adaptation (GDA). GDA covers the major variants as special cases, which allows us to organize them in a comprehensive framework. Moreover, this generalization leads to a new challenging setting where existing methods fail, such as when domain labels are unknown, and class labels are only partially given to each domain. We propose a novel approach to the new setting. The key to our approach is self-supervised class-destructive learning, which enables the learning of class-invariant representations and domain-adversarial classifiers without using any domain labels. Extensive experiments using three benchmark datasets demonstrate that our method outperforms the state-of-the-art UDA methods in the new setting and that it is competitive in existing UDA variations as well.

Generalization of machine learning models trained on a set of source domains on unseen target domains with different statistics, is a challenging problem. While many approaches have been proposed to solve this problem, they only utilize source data during training, but do not take advantage of the fact that a single target example is available at the time of inference. Motivated by this, we propose a method that effectively uses the target sample during inference beyond mere classification. Our method has three components - (i) A label preserving feature or metric transformation on source data such that the source samples are clustered in accordance with their class irrespective of their domain (ii) A generative model trained on the these features (iii) A label-preserving projection of the target point on the source-feature manifold during inference via solving an optimization problem on the input space of the generative model using the learned metric. Finally, the projected target is used in the classifier. Since the projected target feature comes from the source manifold and has the same label as the real target by design, the classifier is expected to perform better on it than the true target. We demonstrate that our method outperforms the state-of-the-art Domain Generalization methods on multiple datasets and tasks.

Domain generalizable (DG) person re-identification (ReID) is a challenging problem because we cannot access any unseen target domain data during training. Almost all the existing DG ReID methods follow the same pipeline where they use a hybrid dataset from multiple source domains for training, and then directly apply the trained model to the unseen target domains for testing. These methods often neglect individual source domains' discriminative characteristics and their relevances w.r.t. the unseen target domains, though both of which can be leveraged to help the model's generalization. To handle the above two issues, we propose a novel method called the relevance-aware mixture of experts (RaMoE), using an effective voting-based mixture mechanism to dynamically leverage source domains' diverse characteristics to improve the model's generalization. Specifically, we propose a decorrelation loss to make the source domain networks (experts) keep the diversity and discriminability of individual domains' characteristics. Besides, we design a voting network to adaptively integrate all the experts' features into the more generalizable aggregated features with domain relevance. Considering the target domains' invisibility during training, we propose a novel learning-to-learn algorithm combined with our relation alignment loss to update the voting network. Extensive experiments demonstrate that our proposed RaMoE outperforms the state-of-the-art methods.

Pedestrian detection is used in many vision based applications ranging from video surveillance to autonomous driving. Despite achieving high performance, it is still largely unknown how well existing detectors generalize to unseen data. This is important because a practical detector should be ready to use in various scenarios in applications. To this end, we conduct a comprehensive study in this paper, using a general principle of direct cross-dataset evaluation. Through this study, we find that existing state-of-the-art pedestrian detectors, though perform quite well when trained and tested on the same dataset, generalize poorly in cross dataset evaluation. We demonstrate that there are two reasons for this trend. Firstly, their designs (e.g. anchor settings) may be biased towards popular benchmarks in the traditional single-dataset training and test pipeline, but as a result largely limit their generalization capability. Secondly, the training source is generally not dense in pedestrians and diverse in scenarios. Under direct cross-dataset evaluation, surprisingly, we find that a general purpose object detector, without pedestrian-tailored adaptation in design, generalizes much better compared to existing state-of-the-art pedestrian detectors. Furthermore, we illustrate that diverse and dense datasets, collected by crawling the web, serve to be an efficient source of pre-training for pedestrian detection. Accordingly, we propose a progressive training pipeline and find that it works well for autonomous-driving oriented pedestrian detection. Consequently, the study conducted in this paper suggests that more emphasis should be put on cross-dataset evaluation for the future design of generalizable pedestrian detectors. Code and models can be accessed at https://github.com/hasanirtiza/Pedestron.

Multi-label image classification is the task of predicting a set of labels corresponding to objects, attributes or other entities present in an image. In this work we propose the Classification Transformer (C-Tran), a general framework for multi-label image classification that leverages Transformers to exploit the complex dependencies among visual features and labels. Our approach consists of a Transformer encoder trained to predict a set of target labels given an input set of masked labels, and visual features from a convolutional neural network. A key ingredient of our method is a label mask training objective that uses a ternary encoding scheme to represent the state of the labels as positive, negative, or unknown during training. Our model shows state-of-the-art performance on challenging datasets such as COCO and Visual Genome. Moreover, because our model explicitly represents the label state during training, it is more general by allowing us to produce improved results for images with partial or extra label annotations during inference. We demonstrate this additional capability in the COCO, Visual Genome, News-500, and CUB image datasets.

In recent years, knowledge distillation has been proved to be an effective solution for model compression. This approach can make lightweight student models acquire the knowledge extracted from cumbersome teacher models. However, previous distillation methods of detection have weak generalization for different detection frameworks and rely heavily on ground truth (GT), ignoring the valuable relation information between instances. Thus, we propose a novel distillation method for detection tasks based on discriminative instances without considering the positive or negative distinguished by GT, which is called general instance distillation (GID). Our approach contains a general instance selection module (GISM) to make full use of feature-based, relation-based and response-based knowledge for distillation. Extensive results demonstrate that the student model achieves significant AP improvement and even outperforms the teacher in various detection frameworks. Specifically, RetinaNet with ResNet-50 achieves 39.1% in mAP with GID on COCO dataset, which surpasses the baseline 36.2% by 2.9%, and even better than the ResNet-101 based teacher model with 38.1% AP.

Recently, a large number of channel attention blocks are proposed to boost the representational power of deep convolutional neural networks (CNNs). These approaches commonly learn the relationship between global contexts and attention activations by using fully-connected layers or linear transformations. However, we empirically find that though many parameters are introduced, these attention blocks may not learn the relationship well. In this paper, we hypothesize that the relationship is predetermined. Based on this hypothesis, we propose a simple yet extremely efficient channel attention block, called Gaussian Context Transformer (GCT), which achieves contextual feature excitation using a Gaussian function that satisfies the presupposed relationship. According to whether the standard deviation of the Gaussian function is learnable, we develop two versions of GCT: GCT-B0 and GCT-B1. GCT-B0 is a parameter-free channel attention block by fixing the standard deviation. It directly maps global contexts to attention activations without learning. In contrast, GCT-B1 is a parameterized channel attention block, which adaptively learns the standard deviation to enhance the mapping ability. Extensive experiments on ImageNet and MS COCO benchmarks demonstrate that our GCTs lead to consistent improvements across various deep CNNs and detectors. Compared with a bank of state-of-the-art channel attention blocks, such as SE and ECA , our GCTs are superior in effectiveness and efficiency.

Video deblurring remains a challenging task due to the complexity of spatially and temporally varying blur. Most of the existing works depend on implicit or explicit alignment for temporal information fusion, which either increases the computational cost or results in suboptimal performance due to misalignment. In this work, we investigate two key factors responsible for deblurring quality: how to fuse spatio-temporal information and from where to collect it. We propose a factorized gated spatio-temporal attention module to perform non-local operations across space and time to fully utilize the available information without depending on alignment. First, we perform spatial aggregation followed by a temporal aggregation step. Next, we adaptively distribute the global spatio-temporal information to each pixel. It shows superior performance compared to existing non-local fusion techniques while being considerably more efficient. To complement the attention module, we propose a reinforcement learning-based framework for selecting keyframes from the neighborhood with the most complementary and useful information. Moreover, our adaptive approach can increase or decrease the frame usage at inference time, depending on the user's need. Extensive experiments on multiple datasets demonstrate the superiority of our method.

Multiple Object Tracking (MOT) has witnessed remarkable advances in recent years. However, existing studies dominantly request prior knowledge of the tracking target (eg, pedestrians), and hence may not generalize well to unseen categories. In contrast, Generic Multiple Object Tracking (GMOT), which requires little prior information about the target, is largely under-explored. In this paper, we make contributions to boost the study of GMOT in three aspects. First, we construct the first publicly available dense GMOT dataset, dubbed GMOT-40, which contains 40 carefully annotated sequences evenly distributed among 10 object categories. In addition, two tracking protocols are adopted to evaluate different characteristics of tracking algorithms. Second, by noting the lack of devoted tracking algorithms, we have designed a series of baseline GMOT algorithms. Third, we perform thorough evaluations on GMOT-40, involving popular MOT algorithms (with necessary modifications) and the proposed baselines. The GMOT-40 benchmark is publicly available at https://github.com/Spritea/GMOT40.

We show that pre-trained Generative Adversarial Networks (GANs), e.g., StyleGAN, can be used as a latent bank to improve the restoration quality of large-factor image super-resolution (SR). While most existing SR approaches attempt to generate realistic textures through learning with adversarial loss, our method, Generative LatEnt bANk (GLEAN), goes beyond existing practices by directly leveraging rich and diverse priors encapsulated in a pre-trained GAN. But unlike prevalent GAN inversion methods that require expensive image-specific optimization at runtime, our approach only needs a single forward pass to generate the upscaled image. GLEAN can be easily incorporated in a simple encoder-bank-decoder architecture with multi-resolution skip connections. Switching the bank allows the method to deal with images from diverse categories, e.g., cat, building, human face, and car. Images upscaled by GLEAN shows clear improvements in terms of fidelity and texture faithfulness in comparison to existing methods.

In this paper, we aim to recognize materials with combined use of auditory and visual perception. To this end, we construct a new dataset named GLAudio that consists of both the geometry of the object being struck and the sound captured from either modal sound synthesis (for virtual objects) or real measurements (for real objects). Besides global geometries, our dataset also takes local geometries around different hitpoints into consideration. This local information is less explored in existing datasets. We demonstrate that local geometry has a greater impact on the sound than the global geometry and offers more cues in material recognition. To extract features from different modalities and perform proper fusion, we propose a new deep neural network GLAVNet that comprises multiple branches and a well-designed fusion module. Once trained on GLAudio, our GLAVNet provides state-of-the-art performance on material identification and supports fine-grained material categorization.

Deep generative models allow for photorealistic image synthesis at high resolutions. But for many applications, this is not enough: content creation also needs to be controllable. While several recent works investigate how to disentangle underlying factors of variation in the data, most of them operate in 2D and hence ignore that our world is three-dimensional. Further, only few works consider the compositional nature of scenes. Our key hypothesis is that incorporating a compositional 3D scene representation into the generative model leads to more controllable image synthesis. Representing scenes as compositional generative neural feature fields allows us to disentangle one or multiple objects from the background as well as individual objects' shapes and appearances while learning from unstructured and unposed image collections without any additional supervision. Combining this scene representation with a neural rendering pipeline yields a fast and realistic image synthesis model. As evidenced by our experiments, our model is able to disentangle individual objects and allows for translating and rotating them in the scene as well as changing the camera pose.

6D pose estimation from a single RGB image is a fundamental task in computer vision. The current top-performing deep learning-based methods rely on an indirect strategy, i.e., first establishing 2D-3D correspondences between the coordinates in the image plane and object coordinate system, and then applying a variant of the PnP/RANSAC algorithm. However, this two-stage pipeline is not end-to-end trainable, thus is hard to be employed for many tasks requiring differentiable poses. On the other hand, methods based on direct regression are currently inferior to geometry-based methods. In this work, we perform an in-depth investigation on both direct and indirect methods, and propose a simple yet effective Geometry-guided Direct Regression Network (GDR-Net) to learn the 6D pose in an end-to-end manner from dense correspondence-based intermediate geometric representations. Extensive experiments show that our approach remarkably outperforms state-of-the-art methods on LM, LM-O and YCB-V datasets. Code is available at https://git.io/GDR-Net.

We present GATSBI, a generative model that can transform a sequence of raw observations into a structured latent representation that fully captures the spatio-temporal context of the agent's actions. In vision-based decision-making scenarios, an agent faces complex high-dimensional observations where multiple entities interact with each other. The agent requires a good scene representation of the visual observation that discerns essential components that consistently propagates along the time horizon. Our method, GATSBI, utilizes unsupervised scene representation learning to successfully separate an active agent, static background, and passive objects. GATSBI then models the interactions reflecting the causal relationships among decomposed entities and predicts physically plausible future states. Our model generalizes to a variety of environments where different types of robots and objects dynamically interact with each other. GATSBI achieves superior performance on scene decompo-sition and video prediction compared to its state-of-the-artcounterparts, and can be readily applied to sequential deci-sion making of an intelligent agent.