CVPR 2021 论文列表

Localizing the camera in a known indoor environment is a key building block for scene mapping, robot navigation, AR, etc. Recent advances estimate the camera pose via optimization over the 2D/3D-3D correspondences established between the coordinates in 2D/3D camera space and 3D world space. Such a mapping is estimated with either a convolution neural network or a decision tree using only the static input image sequence, which makes these approaches vulnerable to dynamic indoor environments that are quite common yet challenging in the real world. To address the aforementioned issues, in this paper, we propose a novel outlier-aware neural tree which bridges the two worlds, deep learning and decision tree approaches. It builds on three important blocks: (a) a hierarchical space partition over the indoor scene to construct the decision tree; (b) a neural routing function, implemented as a deep classification network, employed for better 3D scene understanding; and (c) an outlier rejection module used to filter out dynamic points during the hierarchical routing process. Our proposed algorithm is evaluated on the RIO-10 benchmark developed for camera relocalization in dynamic indoor environments. It achieves robust neural routing through space partitions and outperforms the state-of-the-art approaches by around 30% on camera pose accuracy, while running comparably fast for evaluation.

Vehicle detection with visual sensors like lidar and camera is one of the critical functions enabling autonomous driving. While they generate fine-grained point clouds or high-resolution images with rich information in good weather conditions, they fail in adverse weather (e.g., fog) where opaque particles distort lights and significantly reduce visibility. Thus, existing methods relying on lidar or camera experience significant performance degradation in rare but critical adverse weather conditions. To remedy this, we resort to exploiting complementary radar, which is less impacted by adverse weather and becomes prevalent on vehicles. In this paper, we present Multimodal Vehicle Detection Network (MVDNet), a two-stage deep fusion detector, which first generates proposals from two sensors and then fuses region-wise features between multimodal sensor streams to improve final detection results. To evaluate MVDNet, we create a procedurally generated training dataset based on the collected raw lidar and radar signals from the open-source Oxford Radar Robotcar. We show that the proposed MVDNet surpasses other state-of-the-art methods, notably in terms of Average Precision (AP), especially in adverse weather conditions. The code and data are available at https://github.com/qiank10/MVDNet.

Analyzing complex scenes with Deep Neural Networks is a challenging task, particularly when images contain multiple objects that partially occlude each other. Existing approaches to image analysis mostly process objects independently and do not take into account the relative occlusion of nearby objects. In this paper, we propose a deep network for multi-object instance segmentation that is robust to occlusion and can be trained from bounding box supervision only. Our work builds on Compositional Networks, which learn a generative model of neural feature activations to locate occluders and to classify objects based on their non-occluded parts. We extend their generative model to include multiple objects and introduce a framework for efficient inference in challenging occlusion scenarios. In particular, we obtain feed-forward predictions of the object classes and their instance and occluder segmentations. We introduce an Occlusion Reasoning Module (ORM) that locates erroneous segmentations and estimates the occlusion order to correct them. The improved segmentation masks are, in turn, integrated into the network in a top-down manner to improve the image classification. Our experiments on the KITTI INStance dataset (KINS) and a synthetic occlusion dataset demonstrate the effectiveness and robustness of our model at multi-object instance segmentation under occlusion. Code is publically available at https://github.com/XD7479/Multi-Object-Occlusion.

We present an algorithm for estimating consistent dense depth maps and camera poses from a monocular video. We integrate a learning-based depth prior, in the form of a convolutional neural network trained for single-image depth estimation, with geometric optimization, to estimate a smooth camera trajectory as well as detailed and stable depth reconstruction. Our algorithm combines two complementary techniques: (1) flexible deformation-splines for low-frequency large-scale alignment and (2) geometry-aware depth filtering for high-frequency alignment of fine depth details. In contrast to prior approaches, our method does not require camera poses as input and achieves robust reconstruction for challenging hand-held cell phone captures that contain a significant amount of noise, shake, motion blur, and rolling shutter deformations. Our method quantitatively outperforms state-of-the-arts on the Sintel benchmark for both depth and pose estimations, and attains favorable qualitative results across diverse wild datasets.

Bayesian neural networks have been widely used in many applications because of the distinctive probabilistic representation framework. Even though Bayesian neural networks have been found more robust to adversarial attacks compared with vanilla neural networks, their ability to deal with adversarial noises in practice is still limited. In this paper, we propose Spectral Expectation Bound Regularization (SEBR) to enhance the robustness of Bayesian neural networks. Our theoretical analysis reveals that training with SEBR improves the robustness to adversarial noises. We also prove that training with SEBR can reduce the epistemic uncertainty of the model and hence it can make the model more confident with the predictions, which verifies the robustness of the model from another point of view. Experiments on multiple Bayesian neural network structures and different adversarial attacks validate the correctness of the theoretical findings and the effectiveness of the proposed approach.

We present a self-supervised learning method to learn audio and video representations. Prior work uses the natural correspondence between audio and video to define a standard cross-modal instance discrimination task, where a model is trained to match representations from the two modalities. However, the standard approach introduces two sources of training noise. First, audio-visual correspondences often produce faulty positives since the audio and video signals can be uninformative of each other. To limit the detrimental impact of faulty positives, we optimize a weighted contrastive learning loss, which down-weighs their contribution to the overall loss. Second, since self-supervised contrastive learning relies on random sampling of negative instances, instances that are semantically similar to the base instance can be used as faulty negatives. To alleviate the impact of faulty negatives, we propose to optimize an instance discrimination loss with a soft target distribution that estimates relationships between instances. We validate our contributions through extensive experiments on action recognition tasks and show that they address the problems of audio-visual instance discrimination and improve transfer learning performance.

Most explanation methods in deep learning map importance estimates for a model's prediction back to the original input space. These "visual" explanations are often insufficient, as the model's actual concept remains elusive. Moreover, without insights into the model's semantic concept, it is difficult --if not impossible-- to intervene on the model's behavior via its explanations, called Explanatory Interactive Learning. Consequently, we propose to intervene on a Neuro-Symbolic scene representation, which allows one to revise the model on the semantic level, e.g. "never focus on the color to make your decision". We compiled a novel con-founded visual scene data set, the CLEVR-Hans data set,capturing complex compositions of different objects. The results of our experiments on CLEVR-Hans demonstrate that our semantic explanations, i.e. compositional explanations at a per-object level, can identify confounders that are not identifiable using "visual" explanations only. More importantly, feedback on this semantic level makes it possible to revise the model from focusing on these factors.

This paper presents a method for riggable 3D face reconstruction from monocular images, which jointly estimates a personalized face rig and per-image parameters including expressions, poses, and illuminations. To achieve this goal, we design an end-to-end trainable network embedded with a differentiable in-network optimization. The network first parameterizes the face rig as a compact latent code with a neural decoder, and then estimates the latent code as well as per-image parameters via a learnable optimization. By estimating a personalized face rig, our method goes beyond static reconstructions and enables downstream applications such as video retargeting. In-network optimization explicitly enforces constraints derived from the first principles, thus introduces additional priors than regression-based methods. Finally, data-driven priors from deep learning are utilized to constrain the ill-posed monocular setting and ease the optimization difficulty. Experiments demonstrate that our method achieves SOTA reconstruction accuracy, reasonable robustness and generalization ability, and supports standard face rig applications.

Video quality assessment for User Generated Content (UGC) is an important topic in both industry and academia. Most existing methods only focus on one aspect of the perceptual quality assessment, such as technical quality or compression artifacts. In this paper, we create a large scale dataset to comprehensively investigate characteristics of generic UGC video quality. Besides the subjective ratings and content labels of the dataset, we also propose a DNN-based framework to thoroughly analyze importance of content, technical quality, and compression level in perceptual quality. Our model is able to provide quality scores as well as human-friendly quality indicators, to bridge the gap between low level video signals to human perceptual quality. Experimental results show that our model achieves state-of-the-art correlation with Mean Opinion Scores (MOS).

Glass surfaces appear everywhere. Their existence can however pose a serious problem to computer vision tasks. Recently, a method is proposed to detect glass surfaces by learning multi-scale contextual information. However, as it is only based on a general context integration operation and does not consider any specific glass surface properties, it gets confused when the images contain objects that are similar to glass surfaces and degenerates in challenging scenes with insufficient contexts. We observe that humans often rely on identifying reflections in order to sense the existence of glass and on locating the boundary in order to determine the extent of the glass. Hence, we propose a model for glass surface detection, which consists of two novel modules: (1) a rich context aggregation module (RCAM) to extract multi-scale boundary features from rich context features for locating glass surface boundaries of different sizes and shapes, and (2) a reflection-based refinement module (RRM) to detect reflection and then incorporate it so as to differentiate glass regions from non-glass regions. In addition, we also propose a challenging dataset consisting of 4,012 glass images with annotations for glass surface detection. Our experiments demonstrate that the proposed model outperforms state-of-the-art methods from relevant fields.

Semantic scene understanding from point clouds is particularly challenging as the points reflect only a sparse set of the underlying 3D geometry. Previous works often convert point cloud into regular grids (e.g. voxels or bird-eye view images), and resort to grid-based convolutions for scene understanding. In this work, we introduce RfD-Net that jointly detects and reconstructs dense object surfaces directly from raw point clouds. Instead of representing scenes with regular grids, our method leverages the sparsity of point cloud data and focuses on predicting shapes that are recognized with high objectness. With this design, we decouple the instance reconstruction into global object localization and local shape prediction. It not only eases the difficulty of learning 2-D manifold surfaces from sparse 3D space, the point clouds in each object proposal convey shape details that support implicit function learning to reconstruct any high-resolution surfaces. Our experiments indicate that instance detection and reconstruction present complementary effects, where the shape prediction head shows consistent effects on improving object detection with modern 3D proposal network backbones. The qualitative and quantitative evaluations further demonstrate that our approach consistently outperforms the state-of-the-arts and improves over 11 of mesh IoU in object reconstruction.

State-of-the-art methods for semantic segmentation are based on deep neural networks that are known to be data-hungry. Region-based active learning has shown to be a promising method for reducing data annotation costs. A key design choice for region-based AL is whether to use regularly-shaped regions (e.g., rectangles) or irregularly-shaped region (e.g., superpixels). In this work, we address this question under realistic, click-based measurement of annotation costs. In particular, we revisit the use of superpixels and demonstrate that the inappropriate choice of cost measure (e.g., the percentage of labeled pixels), may cause the effectiveness of the superpixel-based approach to be under-estimated. We benchmark the superpixel-based approach against the traditional "rectangle+polygon"-based approach with annotation cost measured in clicks, and show that the former outperforms on both Cityscapes and PASCAL VOC. We further propose a class-balanced acquisition function to boost the performance of the superpixel-based approach and demonstrate its effectiveness on the evaluation datasets. Our results strongly argue for the use of superpixel-based AL for semantic segmentation and highlight the importance of using realistic annotation costs in evaluating such methods.

Knowledge Distillation (KD) is a popular technique to transfer knowledge from a teacher model or ensemble to a student model. Its success is generally attributed to the privileged information on similarities/consistency between the class distributions or intermediate feature representations of the teacher model and the student model. However, directly pushing the student model to mimic the probabilities/features of the teacher model to a large extent limits the student model in learning undiscovered knowledge/features. In this paper, we propose a novel inheritance and exploration knowledge distillation framework (IE-KD), in which a student model is split into two parts -- inheritance and exploration. The inheritance part is learned with a similarity loss to transfer the existing learned knowledge from the teacher model to the student model, while the exploration part is encouraged to learn representations different from the inherited ones with a dis-similarity loss. Our IE-KD framework is generic and can be easily combined with existing distillation or mutual learning methods for training deep neural networks. Extensive experiments demonstrate that these two parts can jointly push the student model to learn more diversified and effective representations, and our IE-KD can be a general technique to improve the student network to achieve SOTA performance. Furthermore, by applying our IE-KD to the training of two networks, the performance of both can be improved w.r.t. deep mutual learning.

Cross-modal recipe retrieval has recently gained substantial attention due to the importance of food in people's lives, as well as the availability of vast amounts of digital cooking recipes and food images to train machine learning models. In this work, we revisit existing approaches for cross-modal recipe retrieval and propose a simplified end-to-end model based on well established and high performing encoders for text and images. We introduce a hierarchical recipe Transformer which attentively encodes individual recipe components (titles, ingredients and instructions). Further, we propose a self-supervised loss function computed on top of pairs of individual recipe components, which is able to leverage semantic relationships within recipes, and enables training using both image-recipe and recipe-only samples. We conduct a thorough analysis and ablation studies to validate our design choices. As a result, our proposed method achieves state-of-the-art performance in the cross-modal recipe retrieval task on the Recipe1M dataset. We make code and models publicly available.

Low-light image enhancement plays very important roles in low-level vision areas. Recent works have built a great deal of deep learning models to address this task. However, these approaches mostly rely on significant architecture engineering and suffer from high computational burden. In this paper, we propose a new method, named Retinex-inspired Unrolling with Architecture Search (RUAS), to construct lightweight yet effective enhancement network for low-light images in real-world scenario. Specifically, building upon Retinex rule, RUAS first establishes models to characterize the intrinsic underexposed structure of low-light images and unroll their optimization processes to construct our holistic propagation structure. Then by designing a cooperative reference-free learning strategy to discover low-light prior architectures from a compact search space, RUAS is able to obtain a top-performing image enhancement network, which is with fast speed and requires few computational resources. Extensive experiments verify the superiority of our RUAS framework against recently proposed state-of-the-art methods. The project page is available at http://dutmedia.org/RUAS/.

Heatmap regression has become the most prevalent choice for nowadays human pose estimation methods. The ground-truth heatmaps are usually constructed by covering all skeletal keypoints by 2D gaussian kernels. The standard deviations of these kernels are fixed. However, for bottom-up methods, which need to handle a large variance of human scales and labeling ambiguities, the current practice seems unreasonable. To better cope with these problems, we propose the scale-adaptive heatmap regression (SAHR) method, which can adaptively adjust the standard deviation for each keypoint. In this way, SAHR is more tolerant of various human scales and labeling ambiguities. However, SAHR may aggravate the imbalance between fore-background samples, which potentially hurts the improvement of SAHR. Thus, we further introduce the weight-adaptive heatmap regression (WAHR) to help balance the fore-background samples. Extensive experiments show that SAHR together with WAHR largely improves the accuracy of bottom-up human pose estimation. As a result, we finally outperform the state-of-the-art model by +1.5AP and achieve 72.0 AP on COCO test-dev2017, which is comparable with the performances of most top-down methods. Source codes are available at https://github.com/greatlog/SWAHR-HumanPose.

Extensive research in neural style transfer methods has shown that the correlation between features extracted by a pre-trained VGG network has remarkable ability to capture the visual style of an image. Surprisingly, however, this stylization quality is not robust and often degrades significantly when applied to features from more advanced and lightweight networks, such as those in the ResNet family. By performing extensive experiments with different network architectures, we find that residual connections, which represent the main architectural difference between VGG and ResNet, produce feature maps of small entropy, which are not suitable for style transfer. To improve the robustness of the ResNet architecture, we then propose a simple yet effective solution based on a softmax transformation of the feature activations that enhances their entropy. Experimental results demonstrate that this small magic can greatly improve the quality of stylization results, even for networks with random weights. This suggests that the architecture used for feature extraction is more important than the use of learned weights for the task of style transfer.

Text segmentation is a prerequisite in many real-world text-related tasks, e.g., text style transfer, and scene text removal. However, facing the lack of high-quality datasets and dedicated investigations, this critical prerequisite has been left as an assumption in many works, and has been largely overlooked by current research. To bridge this gap, we proposed TextSeg, a large-scale fine-annotated text dataset with six types of annotations: word- and character-wise bounding polygons, masks, and transcriptions. We also introduce Text Refinement Network (TexRNet), a novel text segmentation approach that adapts to the unique properties of text, e.g. non-convex boundary, diverse texture, etc., which often impose burdens on traditional segmentation models. In our TexRNet, we propose text-specific network designs to address such challenges, including key features pooling and attention-based similarity checking. We also introduce trimap and discriminator losses that show significant improvement in text segmentation. Extensive experiments are carried out on both our TextSeg dataset and other existing datasets. We demonstrate that TexRNet consistently improves text segmentation performance by nearly 2% compared to other state-of-the-art segmentation methods. Our dataset and code can be found at https://github.com/SHI-Labs/Rethinking-Text-Segmentation.

There have been many successful implementations of neural style transfer in recent years. In most of these works, the stylization process is confined to the pixel domain. However, we argue that this representation is unnatural because paintings usually consist of brushstrokes rather than pixels. We propose a method to stylize images by optimizing parameterized brushstrokes instead of pixels and further introduce a simple differentiable rendering mechanism. Our approach significantly improves visual quality and enables additional control over the stylization process such as controlling the flow of brushstrokes through user input. We provide qualitative and quantitative evaluations that show the efficacy of the proposed parameterized representation.