ICCV 2021 论文列表

We describe a simple pre-training approach for point clouds. It works in three steps: 1. Mask all points occluded in a camera view; 2. Learn an encoder-decoder model to reconstruct the occluded points; 3. Use the encoder weights as initialisation for downstream point cloud tasks. We find that even when we pre-train on a single dataset (ModelNet40), this method improves accuracy across different datasets and encoders, on a wide range of downstream tasks. Specifically, we show that our method outperforms previous pre-training methods in object classification, and both part-based and semantic segmentation tasks. We study the pre-trained features and find that they lead to wide downstream minima, have high transformation invariance, and have activations that are highly correlated with part labels. Code and data are available at https://github.com/hansen7/OcCo

This paper presents a new task, point cloud object co-segmentation, aiming to segment the common 3D objects in a set of point clouds. We formulate this task as an object point sampling problem, and develop two techniques, the mutual attention module and co-contrastive learning, to enable it. The proposed method employs two point samplers based on deep neural networks, the object sampler and the background sampler. The former targets at sampling points of common objects while the latter focuses on the rest. The mutual attention module explores point-wise correlation across point clouds. It is embedded in both samplers and can identify points with strong cross-cloud correlation from the rest. After extracting features for points selected by the two samplers, we optimize the networks by developing the co-contrastive loss, which minimizes feature discrepancy of the estimated object points while maximizing feature separation between the estimated object and background points. Our method works on point clouds of an arbitrary object class. It is end-to-end trainable and does not need point-level annotations. It is evaluated on the ScanObjectNN and S3DIS datasets and achieves promising results.

Many computer vision problems face difficulties when imaging through turbulent refractive media (e.g., air and water) due to the refraction and scattering of light. These effects cause geometric distortion that requires either handcrafted physical priors or supervised learning methods to remove. In this paper, we present a novel unsupervised network to recover the latent distortion-free image. The key idea is to model non-rigid distortions as deformable grids. Our network consists of a grid deformer that estimates the distortion field and an image generator that outputs the distortion-free image. By leveraging the positional encoding operator, we can simplify the network structure while maintaining fine spatial details in the recovered images. Our method doesn't need to be trained on labeled data and has good transferability across various turbulent image datasets with different types of distortions. Extensive experiments on both simulated and real-captured turbulent images demonstrate that our method can remove both air and water distortions without much customization.

Learning to generate 3D point clouds without 3D supervision is an important but challenging problem. Current solutions leverage various differentiable renderers to project the generated 3D point clouds onto a 2D image plane, and train deep neural networks using the per-pixel difference with 2D ground truth images. However, these solutions are still struggling to fully recover fine structures of 3D shapes, such as thin tubes or planes. To resolve this issue, we propose an unsupervised approach for 3D point cloud generation with fine structures. Specifically, we cast 3D point cloud learning as a 2D projection matching problem. Rather than using entire 2D silhouette images as a regular pixel supervision, we introduce structure adaptive sampling to randomly sample 2D points within the silhouettes as an irregular point supervision, which alleviates the consistency issue of sampling from different view angles. Our method pushes the neural network to generate a 3D point cloud whose 2D projections match the irregular point supervision from different view angles. Our 2D projection matching approach enables the neural network to learn more accurate structure information than using the per-pixel difference, especially for fine and thin 3D structures. Our method can recover fine 3D structures from 2D silhouette images at different resolutions, and is robust to different sampling methods and point number in irregular point supervision. Our method outperforms others under widely used benchmarks. Our code, data and models are available at https://github.com/chenchao15/2D_projection_matching.

We present an unsupervised learning framework for decomposing images into layers of automatically discovered object models. Contrary to recent approaches that model image layers with autoencoder networks, we represent them as explicit transformations of a small set of prototypical images. Our model has three main components: (i) a set of object prototypes in the form of learnable images with a transparency channel, which we refer to as sprites; (ii) differentiable parametric functions predicting occlusions and transformation parameters necessary to instantiate the sprites in a given image; (iii) a layered image formation model with occlusion for compositing these instances into complete images including background. By jointly learning the sprites and occlusion/transformation predictors to reconstruct images, our approach not only yields accurate layered image decompositions, but also identifies object categories and instance parameters. We first validate our approach by providing results on par with the state of the art on standard multi-object synthetic benchmarks (Tetrominoes, Multi-dSprites, CLEVR6). We then demonstrate the applicability of our model to real images in tasks that include clustering (SVHN, GTSRB), cosegmentation (Weizmann Horse) and object discovery from unfiltered social network images. To the best of our knowledge, our approach is the first layered image decomposition algorithm that learns an explicit and shared concept of object type, and is robust enough to be applied to real images.

Image generation has been heavily investigated in computer vision, where one core research challenge is to generate images from arbitrarily complex distributions with little supervision. Generative Adversarial Networks (GANs) as an implicit approach have achieved great successes in this direction and therefore been employed widely. However, GANs are known to suffer from issues such as mode collapse, non-structured latent space, being unable to compute likelihoods, etc. In this paper, we propose a new unsupervised non-parametric method named mixture of infinite conditional GANs or MIC-GANs, to tackle several GAN issues together, aiming for image generation with parsimonious prior knowledge. Through comprehensive evaluations across different datasets, we show that MIC-GANs are effective in structuring the latent space and avoiding mode collapse, and outperform state-of-the-art methods. MICGANs are adaptive, versatile, and robust. They offer a promising solution to several well-known GAN issues. Code available: github.com/yinghdb/MICGANs.

We present MetaUVFS as the first Unsupervised Meta-learning algorithm for Video Few-Shot action recognition. MetaUVFS leverages over 550K unlabeled videos to train a two-stream 2D and 3D CNN architecture via contrastive learning to capture the appearance-specific spatial and action-specific spatio-temporal video features respectively. MetaUVFS comprises a novel Action-Appearance Aligned Meta-adaptation (A3M) module that learns to focus on the action-oriented video features in relation to the appearance features via explicit few-shot episodic meta-learning over unsupervised hard-mined episodes. Our action-appearance alignment and explicit few-shot learner conditions the unsupervised training to mimic the downstream few-shot task, enabling MetaUVFS to significantly outperform all unsupervised methods on few-shot benchmarks. Moreover, unlike previous few-shot action recognition methods that are supervised, MetaUVFS needs neither base-class labels nor a supervised pretrained backbone. Thus, we need to train MetaUVFS just once to perform competitively or sometimes even outperform state-of-the-art supervised methods on popular HMDB51, UCF101, and Kinetics100 few-shot datasets.

Deep learning-based 3D object detection has achieved unprecedented success with the advent of large-scale autonomous driving datasets. However, drastic performance degradation remains a critical challenge for cross-domain deployment. In addition, existing 3D domain adaptive detection methods often assume prior access to the target domain annotations, which is rarely feasible in the real world. To address this challenge, we study a more realistic setting, unsupervised 3D domain adaptive detection, which only utilizes source domain annotations. 1) We first comprehensively investigate the major underlying factors of the domain gap in 3D detection. Our key insight is that geometric mismatch is the key factor of domain shift. 2) Then, we propose a novel and unified framework, Multi-Level Consistency Network (MLC-Net), which employs a teacher-student paradigm to generate adaptive and reliable pseudo-targets. MLC-Net exploits point-, instance- and neural statistics-level consistency to facilitate cross-domain transfer. Extensive experiments demonstrate that MLC-Net outperforms existing state-of-the-art methods (including those using additional target domain information) on standard benchmarks. Notably, our approach is detector-agnostic, which achieves consistent gains on both single- and two-stage 3D detectors. Code will be released.

We propose a deep neural network architecture to infer dense depth from an image and a sparse point cloud. It is trained using a video stream and corresponding synchronized sparse point cloud, as obtained from a LIDAR or other range sensor, along with the intrinsic calibration parameters of the camera. At inference time, the calibration of the camera, which can be different than the one used for training, is fed as an input to the network along with the sparse point cloud and a single image. A Calibrated Backprojection Layer backprojects each pixel in the image to three-dimensional space using the calibration matrix and a depth feature descriptor. The resulting 3D positional encoding is concatenated with the image descriptor and the previous layer output to yield the input to the next layer of the encoder. A decoder, exploiting skip-connections, produces a dense depth map. The resulting Calibrated Backprojection Network, or KBNet, is trained without supervision by minimizing the photometric reprojection error. KBNet imputes missing depth value based on the training set, rather than on generic regularization. We test KBNet on public depth completion benchmarks, where it outperforms the state of the art by 30% indoor and 8% outdoor when the same camera is used for training and testing. When the test camera is different, the improvement reaches 62%.

Shape correspondence from 3D deformation learning has attracted appealing academy interests recently. Nevertheless, current deep learning based methods require the supervision of dense annotations to learn per-point translations, which severely over-parameterize the deformation process. Moreover, they fail to capture local geometric details of original shape via global feature embedding. To address these challenges, we develop a new Unsupervised Dense Deformation Embedding Network (i.e., UD2E-Net), which learns to predict deformations between non-rigid shapes from dense local features. Since it is non-trivial to match deformation-variant local features for deformation prediction, we develop an Extrinsic-Intrinsic Autoencoder to frst encode extrinsic geometric features from source into intrinsic coordinates in a shared canonical shape, with which the decoder then synthesizes corresponding target features. Moreover, a bounded maximum mean discrepancy loss is developed to mitigate the distribution divergence between the synthesized and original features. To learn natural deformation without dense supervision, we introduce a coarse parameterized deformation graph, for which a novel trace and propagation algorithm is proposed to improve both the quality and effciency of the deformation. Our UD2E-Net outperforms state-of-the-art unsupervised methods by 24% on Faust Inter challenge and even supervised methods by 13% on Faust Intra challenge.

Deep convolutional neural networks (CNNs) for video denoising are typically trained with supervision, assuming the availability of clean videos. However, in many applications, such as microscopy, noiseless videos are not available. To address this, we propose an Unsupervised Deep Video Denoiser (UDVD), a CNN architecture designed to be trained exclusively with noisy data. The performance of UDVD is comparable to the supervised state-of-the-art, even when trained only on a single short noisy video. We demonstrate the promise of our approach in real-world imaging applications by denoising raw video, fluorescence-microscopy and electron-microscopy data. In contrast to many current approaches to video denoising, UDVD does not require explicit motion compensation. This is advantageous because motion compensation is computationally expensive, and can be unreliable when the input data are noisy. A gradient-based analysis reveals that UDVD automatically tracks the motion of objects in the input noisy videos. Thus, the network learns to perform implicit motion compensation, even though it is only trained for denoising.

During the last years, convolutional neural networks (CNNs) have triumphed over video quality assessment (VQA) tasks. However, CNN-based approaches heavily rely on annotated data which are typically not available in VQA, leading to the difficulty of model generalization. Recent advances in domain adaptation technique makes it possible to adapt models trained on source data to unlabeled target data. However, due to the distortion diversity and content variation of the collected videos, the intrinsic subjectivity of VQA tasks hampers the adaptation performance. In this work, we propose a curriculum-style unsupervised domain adaptation to handle the cross-domain no-reference VQA problem. The proposed approach could be divided into two stages. In the first stage, we conduct an adaptation between source and target domains to predict the rating distribution for target samples, which can better reveal the subjective nature of VQA. From this adaptation, we split the data in target domain into confident and uncertain subdomains using the proposed uncertainty-based ranking function, through measuring their prediction confidences. In the second stage, by regarding samples in confident subdomain as the easy tasks in the curriculum, a fine-level adaptation is conducted between two subdomains to fine-tune the prediction model. Extensive experimental results on benchmark datasets highlight the superiority of the proposed method over the competing methods in both accuracy and speed. The source code is released at https://github.com/cpf0079/UCDA.

Dance experts often view dance as a hierarchy of information, spanning low-level (raw images, image sequences), mid-levels (human poses and bodypart movements), and high-level (dance genre). We propose a Hierarchical Dance Video Recognition framework (HDVR). HDVR estimates 2D pose sequences, tracks dancers, and then simultaneously estimates corresponding 3D poses and 3D-to-2D imaging parameters, without requiring ground truth for 3D poses. Unlike most methods that work on a single person, our tracking works on multiple dancers, under occlusions. From the estimated 3D pose sequence, HDVR extracts body part movements, and therefrom dance genre. The resulting hierarchical dance representation is explainable to experts. To overcome noise and interframe correspondence ambiguities, we enforce spatial and temporal motion smoothness and photometric continuity over time. We use an LSTM network to extract 3D movement subsequences from which we recognize dance genre. For experiments, we have identified 154 movement types, of 16 body parts, and assembled a new University of Illinois Dance (UID) Dataset, containing 1143 video clips of 9 genres covering 30 hours, annotated with movement and genre labels. Our experimental results demonstrate that our algorithms outperform the state-of-the-art 3D pose estimation methods, which also enhances our dance recognition performance.

Generalization beyond the training distribution is a core challenge in machine learning. The common practice of mixing and shuffling examples when training neural networks may not be optimal in this regard. We show that partitioning the data into well-chosen, non-i.i.d. subsets treated as multiple training environments can guide the learning of models with better out-of-distribution generalization. We describe a training procedure to capture the patterns that are stable across environments while discarding spurious ones. The method makes a step beyond correlation-based learning: the choice of the partitioning allows injecting information about the task that cannot be otherwise recovered from the joint distribution of the training data. We demonstrate multiple use cases with the task of visual question answering, which is notorious for dataset biases. We obtain significant improvements on VQA-CP, using environments built from prior knowledge, existing meta data, or unsupervised clustering. We also get improvements on GQA using annotations of "equivalent questions", and on multi-dataset training (VQA v2 / Visual Genome) by treating them as distinct environments.

High dynamic range (HDR) photography is becoming increasingly popular and available by DSLR and mobile-phone cameras. While deep neural networks (DNN) have greatly impacted other domains of image manipulation, their use for HDR tone-mapping is limited due to the lack of a definite notion of ground-truth solution, which is needed for producing training data. In this paper we describe a new tone-mapping approach guided by the distinct goal of producing low dynamic range (LDR) renditions that best reproduce the visual characteristics of native LDR images. This goal enables the use of an unpaired adversarial training based on unrelated sets of HDR and LDR images, both of which are widely available and easy to acquire. In order to achieve an effective training under this minimal requirements, we introduce the following new steps and components: (i) a range-normalizing pre-process which estimates and applies a different level of curve-based compression, (ii) a loss that preserves the input content while allowing the network to achieve its goal, and (iii) the use of a more concise discriminator network, designed to promote the reproduction of low-level attributes native LDR possess. Evaluation of the resulting network demonstrates its ability to produce photo-realistic artifact-free tone-mapped images, and state-of-the-art performance on different image fidelity indices and visual distances.

We present a simple yet effective unpaired learning based image rain removal method from an unpaired set of synthetic images and real rainy images by exploring the properties of rain maps. The proposed algorithm mainly consists of a semi-supervised learning part and a knowledge distillation part. The semi-supervised part estimates the rain map and reconstructs the derained image based on the well-established layer separation principle. To facilitate rain removal, we develop a rain direction regularizer to constrain the rain estimation network in the semi-supervised learning part. With the estimated rain maps from the semi-supervised learning part, we first synthesize a new paired set by adding to rain-free images based on the superimposition model. The real rainy images and the derained results constitute another paired set. Then we develop an effective knowledge distillation method to explore such two paired sets so that the deraining model in the semi-supervised learning part is distilled. We propose two new rainy datasets, named RainDirection and Real3000, to validate the effectiveness of the proposed method. Both quantitative and qualitative experimental results demonstrate that the proposed method achieves favorable results against state-of-the-art methods in benchmark datasets and real-world images.

Weakly supervised semantic segmentation (WSSS) using image-level classification labels usually utilizes the Class Activation Maps (CAMs) to localize objects of interest in images. While pointing out that CAMs only highlight the most discriminative regions of the classes of interest, adversarial erasing (AE) methods have been proposed to further explore the less discriminative regions. In this paper, we review the potential of the pre-trained classifier which is trained on the raw images. We experimentally verify that the ordinary classifier already has the capability to activate the less discriminative regions if the most discriminative regions are erased to some extent. Based on that, we propose a class-specific AE-based framework that fully exploits the potential of an ordinary classifier. Our framework (1) adopts the ordinary classifier to notify the regions to be erased and (2) generates a class-specific mask for erasing by randomly sampling a single specific class to be erased (target class) among the existing classes on the image for obtaining more precise CAMs. Specifically, with the guidance of the ordinary classifier, the proposed CAMs Generation Network (CGNet) is enforced to generate a CAM of the target class while constraining the CAM not to intrude the object regions of the other classes. Along with the pseudo-labels refined from our CAMs, we achieve the state-of-the-art WSSS performance on both PASCAL VOC 2012 and MS-COCO dataset only with image-level supervision. The code is available at https://github.com/KAIST-vilab/OC-CSE.

Understanding complex social interactions among agents is a key challenge for trajectory prediction. Most existing methods consider the interactions between pairwise traffic agents or in a local area, while the nature of interactions is unlimited, involving an uncertain number of agents and non-local areas simultaneously. Besides, they treat heterogeneous traffic agents the same, namely those among agents of different categories, while neglecting people's diverse reaction patterns toward traffic agents in different categories. To address these problems, we propose a simple yet effective Unlimited Neighborhood Interaction Network (UNIN), which predicts trajectories of heterogeneous agents in multiple categories. Specifically, the proposed unlimited neighborhood interaction module generates the fused-features of all agents involved in an interaction simultaneously, which is adaptive to any number of agents and any range of interaction area. Meanwhile, a hierarchical graph attention module is proposed to obtain category-to-category interaction and agent-to-agent interaction. Finally, parameters of a Gaussian Mixture Model are estimated for generating the future trajectories. Extensive experimental results on benchmark datasets demonstrate a significant performance improvement of our method over the state-of-the-art methods.

Few-shot object detection (FSOD) aims to strengthen the performance of novel object detection with few labeled samples. To alleviate the constraint of few samples, enhancing the generalization ability of learned features for novel objects plays a key role. Thus, the feature learning process of FSOD should focus more on intrinsical object characteristics, which are invariant under different visual changes and therefore are helpful for feature generalization. Unlike previous attempts of the meta-learning paradigm, in this paper, we explore how to enhance object features with intrinsical characteristics that are universal across different object categories. We propose a new prototype, namely universal prototype, that is learned from all object categories. Besides the advantage of characterizing invariant characteristics, the universal prototypes alleviate the impact of unbalanced object categories. After enhancing object features with the universal prototypes, we impose a consistency loss to maximize the agreement between the enhanced features and the original ones, which is beneficial for learning invariant object characteristics. Thus, we develop a new framework of few-shot object detection with universal prototypes ( FSOD ^ up ) that owns the merit of feature generalization towards novel objects. Experimental results on PASCAL VOC and MS COCO show the effectiveness of FSOD ^ up . Particularly, for the 1-shot case of VOC Split2, FSOD ^ up outperforms the baseline by 6.8% in terms of mAP.