CVPR 2021 论文列表

In this paper, we propose a Point-Voxel Recurrent All-Pairs Field Transforms (PV-RAFT) method to estimate scene flow from point clouds. Since point clouds are irregular and unordered, it is challenging to efficiently extract features from all-pairs fields in the 3D space, where all-pairs correlations play important roles in scene flow estimation. To tackle this problem, we present point-voxel correlation fields, which capture both local and long-range dependencies of point pairs. To capture point-based correlations, we adopt the K-Nearest Neighbors search that preserves fine-grained information in the local region. By voxelizing point clouds in a multi-scale manner, we construct pyramid correlation voxels to model long-range correspondences. Integrating these two types of correlations, our PV-RAFT makes use of all-pairs relations to handle both small and large displacements. We evaluate the proposed method on the FlyingThings3D and KITTI Scene Flow 2015 datasets. Experimental results show that PV-RAFT outperforms state-of-the-art methods by remarkable margins.

The effectiveness of learning-based point cloud upsampling pipelines heavily relies on the upsampling modules and feature extractors used therein. For the point upsampling module, we propose a novel model called NodeShuffle, which uses a Graph Convolutional Network (GCN) to better encode local point information from point neighborhoods. NodeShuffle is versatile and can be incorporated into any point cloud upsampling pipeline. Extensive experiments show how NodeShuffle consistently improves state-of-the-art upsampling methods. For feature extraction, we also propose a new multi-scale point feature extractor, called Inception DenseGCN. By aggregating features at multiple scales, this feature extractor enables further performance gain in the final upsampled point clouds. We combine Inception DenseGCN with NodeShuffle into a new point upsampling pipeline called PU-GCN. PU-GCN sets new state-of-art performance with much fewer parameters and more efficient inference.

Non-maximum Suppression (NMS) is an essential post-processing step in modern convolutional neural networks for object detection. Unlike convolutions which are inherently parallel, the de-facto standard for NMS, namely GreedyNMS, cannot be easily parallelized and thus could be the performance bottleneck in convolutional object detection pipelines. MaxpoolNMS is introduced as a parallelizable alternative to GreedyNMS, which in turn enables faster speed than GreedyNMS at comparable accuracy. However, MaxpoolNMS is only capable of replacing the GreedyNMS at the first stage of two-stage detectors like Faster-RCNN. There is a significant drop in accuracy when applying MaxpoolNMS at the final detection stage, due to the fact that MaxpoolNMS fails to approximate GreedyNMS precisely in terms of bounding box selection. In this paper, we propose a general, parallelizable and configurable approach PSRR-MaxpoolNMS, to completely replace GreedyNMS at all stages in all detectors. By introducing a simple Relationship Recovery module and a Pyramid Shifted MaxpoolNMS module, our PSRR-MaxpoolNMS is able to approximate GreedyNMS more precisely than MaxpoolNMS. Comprehensive experiments show that our approach outperforms MaxpoolNMS by a large margin, and it is proven faster than GreedyNMS with comparable accuracy. For the first time, PSRR-MaxpoolNMS provides a fully parallelizable solution for customized hardware design, which can be reused for accelerating NMS everywhere.

Deep learning-based methods have achieved remarkable performance for image dehazing. However, previous studies are mostly focused on training models with synthetic hazy images, which incurs performance drop when the models are used for real-world hazy images. We propose a Principled Synthetic-to-real Dehazing (PSD) framework to improve the generalization performance of dehazing. Starting from a dehazing model backbone that is pre-trained on synthetic data, PSD exploits real hazy images to fine-tune the model in an unsupervised fashion. For the fine-tuning, we leverage several well-grounded physical priors and combine them into a prior loss committee. PSD allows for most of the existing dehazing models as its backbone, and the combination of multiple physical priors boosts dehazing significantly. Through extensive experiments, we demonstrate that our PSD framework establishes the new state-of-the-art performance for real-world dehazing, in terms of visual quality assessed by no-reference quality metrics as well as subjective evaluation and downstream task performance indicator.

Perceptual organization remains one of the very few established theories on the human visual system. It underpinned many pre-deep seminal works on segmentation and detection, yet research has seen a rapid decline since the preferential shift to learning deep models. Of the limited attempts, most aimed at interpreting complex visual scenes using perceptual organizational rules. This has however been proven to be sub-optimal, since models were unable to effectively capture the visual complexity in real-world imagery. In this paper, we rejuvenate the study of perceptual organization, by advocating two positional changes: (i) we examine purposefully generated synthetic data, instead of complex real imagery, and (ii) we ask machines to synthesize novel perceptually-valid patterns, instead of explaining existing data. Our overall answer lies with the introduction of a novel visual challenge -- the challenge of perceptual question answering (PQA). Upon observing example perceptual question-answer pairs, the goal for PQA is to solve similar questions by generating answers entirely from scratch (see Figure 1). Our first contribution is therefore the first dataset of perceptual question-answer pairs, each generated specifically for a particular Gestalt principle. We then borrow insights from human psychology to design an agent that casts perceptual organization as a self-attention problem, where a proposed grid-to-grid mapping network directly generates answer patterns from scratch. Experiments show our agent to outperform a selection of naive and strong baselines. A human study however indicates that ours uses astronomically more data to learn when compared to an average human, necessitating future research (with or without our dataset).

Different from general photo retouching tasks, portrait photo retouching (PPR), which aims to enhance the visual quality of a collection of flat-looking portrait photos, has its special and practical requirements such as human-region priority (HRP) and group-level consistency (GLC). HRP requires that more attention should be paid to human regions, while GLC requires that a group of portrait photos should be retouched to a consistent tone. Models trained on existing general photo retouching datasets, however, can hardly meet these requirements of PPR. To facilitate the research on this high-frequency task, we construct a large-scale PPR dataset, namely PPR10K, which is the first of its kind to our best knowledge. PPR10K contains 1, 681 groups and 11, 161 high-quality raw portrait photos in total. High-resolution segmentation masks of human regions are provided. Each raw photo is retouched by three experts, while they elaborately adjust each group of photos to have consistent tones. We define a set of objective measures to evaluate the performance of PPR and propose strategies to learn PPR models with good HRP and GLC performance. The constructed PPR10K dataset provides a good benchmark for studying automatic PPR methods, and experiments demonstrate that the proposed learning strategies are effective to improve the retouching performance. Datasets and codes are available: https://github.com/csjliang/PPR10K.

We propose POse-guided SElective Fusion (POSEFusion), a single-view human volumetric capture method that leverages tracking-based methods and tracking-free inference to achieve high-fidelity and dynamic 3D reconstruction. By contributing a novel reconstruction framework which contains pose-guided keyframe selection and robust implicit surface fusion, our method fully utilizes the advantages of both tracking-based methods and tracking-free inference methods, and finally enables the high-fidelity reconstruction of dynamic surface details even in the invisible regions. We formulate the keyframe selection as a dynamic programming problem to guarantee the temporal continuity of the reconstructed sequence. Moreover, the novel robust implicit surface fusion involves an adaptive blending weight to preserve high-fidelity surface details and an automatic collision handling method to deal with the potential self-collisions. Overall, our method enables high-fidelity and dynamic capture in both visible and invisible regions from a single RGBD camera, and the results and experiments show that our method outperforms state-of-the-art methods.

The task of point cloud completion aims to predict the missing part for an incomplete 3D shape. A widely used strategy is to generate a complete point cloud from the incomplete one. However, the unordered nature of point clouds will degrade the generation of high-quality 3D shapes, as the detailed topology and structure of discrete points are hard to be captured by the generative process only using a latent code. In this paper, we address the above problem by reconsidering the completion task from a new perspective, where we formulate the prediction as a point cloud deformation process. Specifically, we design a novel neural network, named PMP-Net, to mimic the behavior of an earth mover. It moves move each point of the incomplete input to complete the point cloud, where the total distance of point moving paths (PMP) should be shortest. Therefore, PMP-Net predicts a unique point moving path for each point according to the constraint of total point moving distances. As a result, the network learns a strict and unique correspondence on point-level, and thus improves the quality of the predicted complete shape. We conduct comprehensive experiments on Completion3D and PCN datasets, which demonstrate our advantages over the state-of-the-art point cloud completion methods. Code will be available at https://github.com/diviswen/PMP-Net.

In this paper, we propose a progressive margin loss (PML) approach for unconstrained facial age classification. Conventional methods make strong assumption on that each class owns adequate instances to outline its data distribution, likely leading to bias prediction where the training samples are sparse across age classes. Instead, our PML aims to adaptively refine the age label pattern by enforcing a couple of margins, which fully takes in the in-between discrepancy of the intra-class variance, inter-class variance and class-center. Our PML typically incorporates with the ordinal margin and the variational margin, simultaneously plugging in the globally-tuned deep neural network paradigm. More specifically, the ordinal margin learns to exploit the correlated relationship of the real-world age labels. Accordingly, the variational margin is leveraged to minimize the influence of head classes that misleads the prediction of tailed samples. Moreover, our optimization carefully seeks a series of indicator curricula to achieve robust and efficient model training. Extensive experimental results on three face aging datasets demonstrate that our PML achieves compelling performance compared to state of the arts. Code will be made publicly.

Deep learning approaches are nowadays ubiquitously used to tackle computer vision tasks such as semantic segmentation, requiring large datasets and substantial computational power. Continual learning for semantic segmentation (CSS) is an emerging trend that consists in updating an old model by sequentially adding new classes. However, continual learning methods are usually prone to catastrophic forgetting. This issue is further aggravated in CSS where, at each step, old classes from previous iterations are collapsed into the background. In this paper, we propose Local POD, a multi-scale pooling distillation scheme that preserves long- and short-range spatial relationships at feature level. Furthermore, we design an entropy-based pseudo-labelling of the background w.r.t. classes predicted by the old model to deal with background shift and avoid catastrophic forgetting of the old classes. Our approach, called PLOP, significantly outperforms state-of-the-art methods in existing CSS scenarios, as well as in newly proposed challenging benchmarks.

In this paper, we propose a self-supervised single-view pixel-level accurate depth estimation network, called PLADE-Net. The PLADE-Net is the first work that shows unprecedented accuracy levels, exceeding 95% in terms of the \delta^1 metric on the challenging KITTI dataset. Our PLADE-Net is based on a new network architecture with neural positional encoding and a novel loss function that borrows from the closed-form solution of the matting Laplacian to learn pixel-level accurate depth estimation from stereo images. Neural positional encoding allows our PLADE-Net to obtain more consistent depth estimates by letting the network reason about location-specific image properties such as lens and projection distortions. Our novel distilled matting Laplacian loss allows our network to predict sharp depths at object boundaries and more consistent depths in highly homogeneous regions. Our proposed method outperforms all previous self-supervised single-view depth estimation methods by a large margin on the challenging KITTI dataset, with unprecedented levels of accuracy. Furthermore, our PLADE-Net, naively extended for stereo inputs, outperforms the most recent self-supervised stereo methods, even without any advanced blocks like 1D correlations, 3D convolutions, or spatial pyramid pooling. We present extensive ablation studies and experiments that support our method's effectiveness on the KITTI, CityScapes, and Make3D datasets.

Person image synthesis, e.g., pose transfer, is a challenging problem due to large variation and occlusion. Existing methods have difficulties predicting reasonable invisible regions and fail to decouple the shape and style of clothing, which limits their applications on person image editing. In this paper, we propose PISE, a novel two-stage generative model for person image synthesis and editing, which can generate realistic person images with desired poses, textures, and semantic layouts. To better predict the invisible region, we first synthesize a human parsing map aligned with the target pose to represent the shape of clothing by a parsing generator, and then generate the final image by an image generator. To decouple the shape and style of clothing, we propose joint global and local per-region encoding and normalization to predict the reasonable style of clothing for invisible regions. We also propose spatial-aware normalization to retain the spatial context relationship in the source image. The results of qualitative and quantitative experiments demonstrate the superiority of our model. Besides, the results of texture transfer and parsing editing show that our model can be applied to person image editing.

Convolutional video models have an order of magnitude larger computational complexity than their counterpart image-level models. Constrained by computational resources, there is no model or training method that can train long video sequences end-to-end. Currently, the main-stream method is to split a raw video into clips, leading to incomplete fragmentary temporal information flow. Inspired by natural language processing techniques dealing with long sentences, we propose to treat videos as serial fragments satisfying Markov property, and train it as a whole by progressively propagating information through the temporal dimension in multiple steps. This progressive training (PGT) method is able to train long videos end-to-end with limited resources and ensures the effective transmission of information. As a general and robust training method, we empirically demonstrate that it yields significant performance improvements on different models and datasets. As an illustrative example, the proposed method improves SlowOnly network by 3.7 mAP on Charades and 1.9 top-1 accuracy on Kinetics with negligible parameter and computation overhead. The code is attached in supplementary files and will be published with this paper.

We propose PD-GAN, a probabilistic diverse GAN forimage inpainting. Given an input image with arbitrary holeregions, PD-GAN produces multiple inpainting results withdiverse and visually realistic content. Our PD-GAN is builtupon a vanilla GAN which generates images based on random noise. During image generation, we modulate deepfeatures of input random noise from coarse-to-fine by injecting an initially restored image and the hole regions inmultiple scales. We argue that during hole filling, the pixels near the hole boundary should be more deterministic(i.e., with higher probability trusting the context and initially restored image to create natural inpainting boundary), while those pixels lie in the center of the hole shouldenjoy more degrees of freedom (i.e., more likely to dependon the random noise for enhancing diversity). To this end, we propose spatially probabilistic diversity normalization(SPDNorm) inside the modulation to model the probabilityof generating a pixel conditioned on the context information. SPDNorm dynamically balances the realism and diversity inside the hole region, making the generated content more diverse towards the hole center and resembleneighboring image content more towards the hole boundary. Meanwhile, we propose a perceptual diversity loss tofurther empower PD-GAN for diverse content generation. Experiments on benchmark datasets including CelebA-HQ, Places2 and Paris Street View indicate that PD-GAN is ef-fective for diverse and visually realistic image restoration.

Local processing is an essential feature of CNNs and other neural network architectures -- it is one of the reasons why they work so well on images where relevant information is, to a large extent, local. However, perspective effects stemming from the projection in a conventional camera vary for different global positions in the image. We introduce Perspective Crop Layers (PCLs) -- a form of perspective crop of the region of interest based on the camera geometry -- and show that accounting for the perspective consistently improves the accuracy of state-of-the-art 3D pose reconstruction methods. PCLs are modular neural network layers, which, when inserted into existing CNN and MLP architectures, deterministically remove the location-dependent perspective effects while leaving end-to-end training and the number of parameters of the underlying neural network unchanged. We demonstrate that PCL leads to improved 3D human pose reconstruction accuracy for CNN architectures that use cropping operations, such as spatial transformer networks (STN), and, somewhat surprisingly, MLPs used for 2D-to-3D keypoint lifting. Our conclusion is that it is important to utilize camera calibration information when available, for classical and deep-learning-based computer vision alike. PCL offers an easy way to improve the accuracy of existing 3D reconstruction networks by making them geometry aware. Our code is publicly available at github.com/yu-frank/PerspectiveCropLayers.

Recent success in casting Non-rigid Structure from Motion (NRSfM) as an unsupervised deep learning problem has raised fundamental questions about what novelty in NRSfM prior could the deep learning offer. In this paper we advocate for a 3D deep auto-encoder framework to be used explicitly as the NRSfM prior. The framework is unique as: (i) it learns the 3D auto-encoder weights solely from 2D projected measurements, and (ii) it is Procrustean in that it jointly resolves the unknown rigid pose for each shape instance. We refer to this architecture as a Procustean Autoencoder for Unsupervised Lifting (PAUL), and demonstrate state-of-the-art performance across a number of benchmarks in comparison to recent innovations such as Deep NRSfM and C3PDO.

Anomaly detection methods require high-quality features. In recent years, the anomaly detection community has attempted to obtain better features using advances in deep self-supervised feature learning. Surprisingly, a very promising direction, using pre-trained deep features, has been mostly overlooked. In this paper, we first empirically establish the perhaps expected, but unreported result, that combining pre-trained features with simple anomaly detection and segmentation methods convincingly outperforms, much more complex, state-of-the-art methods. In order to obtain further performance gains in anomaly detection, we adapt pre-trained features to the target distribution. Although transfer learning methods are well established in multi-class classification problems, the one-class classification (OCC) setting is not as well explored. It turns out that naive adaptation methods, which typically work well in supervised learning, often result in catastrophic collapse (feature deterioration) and reduce performance in OCC settings. A popular OCC method, DeepSVDD, advocates using specialized architectures, but this limits the adaptation performance gain. We propose two methods for combating collapse: i) a variant of early stopping that dynamically learns the stopping iteration ii) elastic regularization inspired by continual learning. Our method, PANDA, outperforms the state-of-the-art in the OCC, outlier exposure and anomaly segmentation settings by large margins.

We introduce Position Adaptive Convolution (PAConv), a generic convolution operation for 3D point cloud processing. The key of PAConv is to construct the convolution kernel by dynamically assembling basic weight matrices stored in Weight Bank, where the coefficients of these weight matrices are self-adaptively learned from point positions through ScoreNet. In this way, the kernel is built in a data-driven manner, endowing PAConv with more flexibility than 2D convolutions to better handle the irregular and unordered point cloud data. Besides, the complexity of the learning process is reduced by combining weight matrices instead of brutally predicting kernels from point positions. Furthermore, different from the existing point convolution operators whose network architectures are often heavily engineered, we integrate our PAConv into classical MLP-based point cloud pipelines without changing network configurations. Even built on simple networks, our method still approaches or even surpasses the state-of-the-art models, and significantly improves baseline performance on both classification and segmentation tasks, yet with decent efficiency. Thorough ablation studies and visualizations are provided to understand PAConv. Code is released on https://github.com/CVMI-Lab/PAConv.

Deep neural networks for video classification, just like image classification networks, may be subjected to adversarial manipulation. The main difference between image classifiers and video classifiers is that the latter usually use temporal information contained within the video. In this work we present a manipulation scheme for fooling video classifiers by introducing a flickering temporal perturbation that in some cases may be unnoticeable by human observers and is implementable in the real world. After demonstrating the manipulation of action classification of single videos, we generalize the procedure to make universal adversarial perturbation, achieving high fooling ratio. In addition, we generalize the universal perturbation and produce a temporal-invariant perturbation, which can be applied to the video without synchronizing the perturbation to the input. The attack was implemented on several target models and the transferability of the attack was demonstrated. These properties allow us to bridge the gap between simulated environment and real-world application, as will be demonstrated in this paper for the first time for an over-the-air flickering attack.