ICCV 2021 Papers

Scene understanding under low-light conditions is a challenging problem. This is due to the small number of photons captured by the camera and the resulting low signal-to-noise ratio (SNR). Single-photon cameras (SPCs) are an emerging sensing modality that are capable of capturing images with high sensitivity. Despite having minimal read-noise, images captured by SPCs in photon-starved conditions still suffer from strong shot noise, preventing reliable scene inference. We propose photon scale-space -- a collection of high-SNR images spanning a wide range of photons-per-pixel (PPP) levels (but same scene content) as guides to train inference model on low photon flux images. We develop training techniques that push images with different illumination levels closer to each other in feature representation space. The key idea is that having a spectrum of different brightness levels during training enables effective guidance, and increases robustness to shot noise even in extreme noise cases. Based on the proposed approach, we demonstrate, via simulations and real experiments with a SPAD camera, high-performance on various inference tasks such as image classification and monocular depth estimation under ultra low-light, down to <1 PPP.

Large-scale point cloud semantic segmentation has wide applications. Current popular researches mainly focus on fully supervised learning which demands expensive and tedious manual point-wise annotation. Weakly supervised learning is an alternative way to avoid this exhausting annotation. However, for large-scale point clouds with few labeled points, the network is difficult to extract discriminative features for unlabeled points, as well as the regularization of topology between labeled and unlabeled points is usually ignored, resulting in incorrect segmentation results. To address this problem, we propose a perturbed self-distillation (PSD) framework. Specifically, inspired by self-supervised learning, we construct the perturbed branch and enforce the predictive consistency among the perturbed branch and original branch. In this way, the graph topology of the whole point cloud can be effectively established by the introduced auxiliary supervision, such that the information propagation between the labeled and unlabeled points will be realized. Besides point-level supervision, we present a well-integrated context-aware module to explicitly regularize the affinity correlation of labeled points. Therefore, the graph topology of the point cloud can be further refined. The experimental results evaluated on three large-scale datasets show the large gain (3.0% on average) against recent weakly supervised methods and comparable results to some fully supervised methods.

The popularization of intelligent devices including smartphones and surveillance cameras results in more serious privacy issues. De-identification is regarded as an effective tool for visual privacy protection with the process of concealing or replacing identity information. Most of the existing de-identification methods suffer from some limitations since they mainly focus on the protection process and are usually non-reversible. In this paper, we propose a personalized and invertible de-identification method based on the deep generative model, where the main idea is introducing a user-specific password and an adjustable parameter to control the direction and degree of identity variation. Extensive experiments demonstrate the effectiveness and generalization of our proposed framework for both face de-identification and recovery.

Trajectory prediction is confronted with the dilemma to capture the multi-modal nature of future dynamics with both diversity and accuracy. In this paper, we propose a distribution discrimination method (DisDis) to predict personalized motion pattern by distinguishing the potential distributions in a self-supervised manner. The key motivation of DisDis is the observation that the motion pattern of each person is personalized due to his/her habit, character, or goal. Specifically, we learn the latent distribution to represent different motion patterns and optimize it by contrastive discrimination. The contrastive distribution discrimination encourages latent distributions to be discriminative. Our method could be seamlessly integrated with existing multi-modal stochastic predictive models as a plug-and-play module to learn the more discriminative latent distribution. To evaluate the latent distribution, we further propose a new metric, probability cumulative minimum distance (PCMD) curve, which cumulatively calculates the minimum distance on the sorted probabilities. Experimental results on the ETH and UCY datasets show the effectiveness of our method.

Semantic segmentation models trained on public datasets have achieved great success in recent years. However, these models didn't consider the personalization issue of segmentation though it is important in practice. In this paper, we address the problem of personalized image segmentation. The objective is to generate more accurate segmentation results on unlabeled personalized images by investigating the data's personalized traits. To open up future research in this area, we collect a large dataset containing various users' personalized images called PSS (Personalized Semantic Segmentation). We also survey some recent researches related to this problem and report their performance on our dataset. Furthermore, by observing the correlation among a user's personalized images, we propose a baseline method that incorporates the inter-image context when segmenting certain images. Extensive experiments show that our method outperforms the existing methods on the proposed dataset. The code and the PSS dataset are available at https://mmcheng.net/pss/.

Fine-grained 3D segmentation is an important task in 3D object understanding, especially in applications such as intelligent manufacturing or parts analysis for 3D objects. However, many challenges involved in such problem are yet to be solved, such as i) interpreting the complex structures located in different regions for 3D objects; ii) capturing fine-grained structures with sufficient topology correctness. Current deep learning and graph machine learning methods fail to tackle such challenges and thus provide inferior performance in fine-grained 3D analysis. In this work, methods in topological data analysis are incorporated with geometric deep learning model for the task of fine-grained segmentation for 3D objects. We propose a novel neural network model called Persistent Homology based Graph Convolution Network (PHGCN), which i) integrates persistent homology into graph convolution network to capture multi-scale structural information that can accurately represent complex structures for 3D objects; ii) applies a novel Persistence Diagram Loss that provides sufficient topology correctness for segmentation over the fine-grained structures. Extensive experiments on fine-grained 3D segmentation validate the effectiveness of the proposed PHGCN model and show significant improvements over current state-of-the-art methods.

Deep learning algorithms have made significant progress in dynamic scene deblurring. However, several challenges are still unsettled: 1) The degree and scale of blur in different regions of a blurred image can have a considerable variation in a large range. However, the traditional input pyramid or downscaling-upscaling, is designed to have limited and inflexible perceptual variousness to cope with large blur scale variation. 2) The nonlocal block is proved to be effective in the image enhancement tasks, but it requires high computation and memory cost. In this paper, we are the first to propose a light-weight globally-analyzing module into the image deblurring field, named Light Global Context Refinement (LGCR) module. With exponentially lower cost, it achieves even better performance than the nonlocal unit. Moreover, we propose the Perceptual Variousness Block (PVB) and PVB-piling strategy. By placing PVB repeatedly, the whole method possesses abundant reception field spectrum to be aware of the blur with various degrees and scales. Comprehensive experimental results from the different benchmarks and assessment metrics show that our method achieves excellent performance to set a new state-of-the-art in motion deblurring.

3D LiDAR (light detection and ranging) semantic segmentation is important in scene understanding for many applications, such as auto-driving and robotics. For example, for autonomous cars equipped with RGB cameras and LiDAR, it is crucial to fuse complementary information from different sensors for robust and accurate segmentation. Existing fusion-based methods, however, may not achieve promising performance due to the vast difference between the two modalities. In this work, we investigate a collaborative fusion scheme called perception-aware multi-sensor fusion (PMF) to exploit perceptual information from two modalities, namely, appearance information from RGB images and spatio-depth information from point clouds. To this end, we first project point clouds to the camera coordinates to provide spatio-depth information for RGB images. Then, we propose a two-stream network to extract features from the two modalities, separately, and fuse the features by effective residual-based fusion modules. Moreover, we propose additional perception-aware losses to measure the perceptual difference between the two modalities. Extensive experiments on two benchmark data sets show the superiority of our method. For example, on nuScenes, our PMF outperforms the state-of-the-art method by 0.8% in mIoU.

People navigating in unfamiliar buildings take advantage of myriad visual, spatial and semantic cues to efficiently achieve their navigation goals. Towards equipping computational agents with similar capabilities, we introduce Pathdreamer, a visual world model for agents navigating in novel indoor environments. Given one or more previous visual observations, Pathdreamer generates plausible high-resolution 360deg visual observations (RGB, semantic segmentation and depth) for viewpoints that have not been visited, in buildings not seen during training. In regions of high uncertainty (e.g. predicting around corners, imagining the contents of an unseen room), Pathdreamer can predict diverse scenes, allowing an agent to sample multiple realistic outcomes for a given trajectory. We demonstrate that Pathdreamer encodes useful and accessible visual, spatial and semantic knowledge about human environments by using it in the downstream task of Vision-and-Language Navigation (VLN). Specifically, we show that planning ahead with Pathdreamer brings about half the benefit of looking ahead at actual observations from unobserved parts of the environment. We hope that Pathdreamer will help unlock model-based approaches to challenging embodied navigation tasks such as navigating to specified objects and VLN.

Recent learning-based multi-view stereo (MVS) methods show excellent performance with dense cameras and small depth ranges. However, non-learning based approaches still outperform for scenes with large depth ranges and sparser wide-baseline views, in part due to their PatchMatch optimization over pixelwise estimates of depth, normals, and visibility. In this paper, we propose an end-to-end trainable PatchMatch-based MVS approach that combines advantages of trainable costs and regularizations with pixelwise estimates. To overcome the challenge of the non-differentiable PatchMatch optimization that involves iterative sampling and hard decisions, we use reinforcement learning to minimize expected photometric cost and maximize likelihood of ground truth depth and normals. We incorporate normal estimation by using dilated patch kernels, and propose a recurrent cost regularization that applies beyond frontal plane-sweep algorithms to our pixelwise depth/normal estimates. We evaluate our method on widely used MVS benchmarks, ETH3D and Tanks and Temples (TnT), and compare to other state of the art learning based MVS models. On ETH3D, our method outperforms other recent learning-based approaches and performs comparably on advanced TnT.

3D perception of object shapes from RGB image input is fundamental towards semantic scene understanding, grounding image-based perception in our spatially 3-dimensional real-world environments. To achieve a mapping between image views of objects and 3D shapes, we leverage CAD model priors from existing large-scale databases, and propose a novel approach towards constructing a joint embedding space between 2D images and 3D CAD models in a patch-wise fashion -- establishing correspondences between patches of an image view of an object and patches of CAD geometry. This enables part similarity reasoning for retrieving similar CADs to a new image view without exact matches in the database. Our patch embedding provides more robust CAD retrieval for shape estimation in our end-to-end estimation of CAD model shape and pose for detected objects in a single input image. Experiments on in-the-wild, complex imagery from ScanNet show that our approach is more robust than state of the art in real-world scenarios without any exact CAD matches.

The non-local self-similarity property of natural images has been exploited extensively for solving various image processing problems. When it comes to video sequences, harnessing this force is even more beneficial due to the temporal redundancy. In the context of image and video denoising, many classically-oriented algorithms employ self-similarity, splitting the data into overlapping patches, gathering groups of similar ones and processing these together somehow. With the emergence of convolutional neural networks (CNN), the patch-based framework has been abandoned. Most CNN denoisers operate on the whole image, leveraging non-local relations only implicitly by using a large receptive field. This work proposes a novel approach for leveraging self-similarity in the context of video denoising, while still relying on a regular convolutional architecture. We introduce a concept of patch-craft frames - artificial frames that are similar to the real ones, built by tiling matched patches. Our algorithm augments video sequences with patch-craft frames and feeds them to a CNN. We demonstrate the substantial boost in denoising performance obtained with the proposed approach.

Few-shot Learning has been studied to mimic human visual capabilities and learn effective models without the need of exhaustive human annotation. Even though the idea of meta-learning for adaptation has dominated the few-shot learning methods, how to train a feature extractor is still a challenge. In this paper, we focus on the design of training strategy to obtain an elemental representation such that the prototype of each novel class can be estimated from a few labeled samples. We propose a two-stage training scheme, Partner-Assisted Learning (PAL), which first trains a partner encoder to model pair-wise similarities and extract features serving as soft-anchors, and then trains a main encoder by aligning its outputs with soft-anchors while attempting to maximize classification performance. Two alignment constraints from logit-level and feature-level are designed individually. For each few-shot task, we perform prototype classification. Our method consistently outperforms the state-of-the-art method on four benchmarks. Detailed ablation studies of PAL are provided to justify the selection of each component involved in training.

Partial Domain Adaptation (PDA) is a practical and general domain adaptation scenario, which relaxes the fully shared label space assumption such that the source label space subsumes the target one. The key challenge of PDA is the issue of negative transfer caused by source-only classes. For videos, such negative transfer could be triggered by both spatial and temporal features, which leads to a more challenging Partial Video Domain Adaptation (PVDA) problem. In this paper, we propose a novel Partial Adversarial Temporal Attentive Network (PATAN) to address the PVDA problem by utilizing both spatial and temporal features for filtering source-only classes. Besides, PATAN constructs effective overall temporal features by attending to local temporal features that contribute more toward the class filtration process. We further introduce new benchmarks to facilitate research on PVDA problems, covering a wide range of PVDA scenarios. Empirical results demonstrate the state-of-the-art performance of our proposed PATAN across the multiple PVDA benchmarks.

Human-oriented image captioning with both high diversity and accuracy is a challenging task in vision+language modeling. The reinforcement learning (RL) based frameworks promote the accuracy of image captioning, yet seriously hurt the diversity. In contrast, other methods based on variational auto-encoder (VAE) or generative adversarial network (GAN) can produce diverse yet less accurate captions. In this work, we devote our attention to promote the diversity of RL-based image captioning. To be specific, we devise a partial off-policy learning scheme to balance accuracy and diversity. First, we keep the model exposed to varied candidate captions by sampling from the initial state before RL launched. Second, a novel criterion named max-CIDEr is proposed to serve as the reward for promoting diversity. We combine the above-mentioned off-policy strategy with the on-policy one to moderate the exploration effect, further balancing the diversity and accuracy for human-like image captioning. Experiments show that our method locates the closest to human performance in the diversity-accuracy space, and achieves the highest Pearson correlation as 0.337 with human performance.

This paper tackles the problem of table structure pars-ing (TSP) from images in the wild. In contrast to existingstudies that mainly focus on parsing well-aligned tabularimages with simple layouts from scanned PDF documents,we aim to establish a practical table structure parsing sys-tem for real-world scenarios where tabular input imagesare taken or scanned with severe deformation, bending orocclusions. For designing such a system, we propose anapproach named Cycle-CenterNet on the top of CenterNetwith a novel cycle-pairing module to simultaneously detectand group tabular cells into structured tables. In the cycle-pairing module, a new pairing loss function is proposed forthe network training. Alongside with our Cycle-CenterNet,we also present a large-scale dataset, named Wired Tablein the Wild (WTW), which includes well-annotated structureparsing of multiple style tables in several scenes like photo,scanning files, web pages,etc.. In experiments, we demon-strate that our Cycle-CenterNet consistently achieves thebest accuracy of table structure parsing on the new WTWdataset by 24.6% absolute improvement evaluated by theTEDS metric. A more comprehensive experimental analysisalso validates the advantages of our proposed methods forthe TSP task.

In this paper, we propose Parametric Contrastive Learning (PaCo) to tackle long-tailed recognition. Based on theoretical analysis, we observe supervised contrastive loss tends to bias on high-frequency classes and thus increases the difficulty of imbalanced learning. We introduce a set of parametric class-wise learnable centers to rebalance from an optimization perspective. Further, we analyze our PaCo loss under a balanced setting. Our analysis demonstrates that PaCo can adaptively enhance the intensity of pushing samples of the same class close as more samples are pulled together with their corresponding centers and benefit hard example learning. Experiments on long-tailed CIFAR, ImageNet, Places, and iNaturalist 2018 manifest the new state-of-the-art for long-tailed recognition. On full ImageNet, models trained with PaCo loss surpass supervised contrastive learning across various ResNet backbones, e.g., our ResNet-200 achieves 81.8% top-1 accuracy. Our code is available at https://github.com/dvlab-research/Parametric-Contrastive-Learning.

Object detection has been widely used in many safety-critical tasks, such as autonomous driving. However, its vulnerability to adversarial examples has not been sufficiently studied, especially under the practical scenario of black-box attacks, where the attacker can only access the query feedback of predicted bounding-boxes and top-1 scores returned by the attacked model. Compared with black-box attack to image classification, there are two main challenges in black-box attack to detection. Firstly, even if one bounding-box is successfully attacked, another sub-optimal bounding-box may be detected near the attacked bounding-box. Secondly, there are multiple bounding-boxes, leading to very high attack cost. To address these challenges, we propose a Parallel Rectangle Flip Attack (PRFA) via random search. Specifically, we generate perturbations in each rectangle patch to avoid sub-optimal detection near the attacked region. Besides, utilizing the observation that adversarial perturbations mainly locate around objects' contours and critical points under white-box attacks, the search space of attacked rectangles is reduced to improve the attack efficiency. Moreover, we develop a parallel mechanism of attacking multiple rectangles simultaneously to further accelerate the attack process. Extensive experiments demonstrate that our method can effectively and efficiently attack various popular object detectors, including anchor-based and anchor-free, and generate transferable adversarial examples.

Conventional deep image inpainting methods are based on auto-encoder architecture, in which the spatial details of images will be lost in the down-sampling process, leading to the degradation of generated results. Also, the structure information in deep layers and texture information in shallow layers of the auto-encoder architecture can not be well integrated. Differing from the conventional image inpainting architecture, we design a parallel multi-resolution inpainting network with multi-resolution partial convolution, in which low-resolution branches focus on the global structure while high-resolution branches focus on the local texture details. All these high- and low-resolution streams are in parallel and fused repeatedly with multi-resolution masked representation fusion so that the reconstructed images are semantically robust and textually plausible. Experimental results show that our method can effectively fuse structure and texture information, producing more realistic results than state-of-the-art methods.