ICCV 2021 论文列表

3D deep learning has been increasingly more popular for a variety of tasks including many safety-critical applications. However, recently several works raise the security issues of 3D deep models. Although most of them consider adversarial attacks, we identify that backdoor attack is indeed a more serious threat to 3D deep learning systems but remains unexplored. We present the backdoor attacks in 3D point cloud with a unified framework that exploits the unique properties of 3D data and networks. In particular, we design two attack approaches on point cloud: the poison-label backdoor attack (PointPBA) and the clean-label backdoor attack (PointCBA). The first one is straightforward and effective in practice, while the latter is more sophisticated assuming there are certain data inspections. The attack algorithms are mainly motivated and developed by 1) the recent discovery of 3D adversarial samples suggesting the vulnerability of deep models under spatial transformation; 2) the proposed feature disentanglement technique that manipulates the feature of the data through optimization methods and its potential to embed a new task. Extensive experiments show the efficacy of the PointPBA with over 95% success rate across various 3D datasets and models, and the more stealthy PointCBA with around 50% success rate. Our proposed backdoor attack in 3D point cloud is expected to perform as a baseline for improving the robustness of 3D deep models.

Learning an effective representation of 3D point clouds requires a good metric to measure the discrepancy between two 3D point sets, which is non-trivial due to their irregularity. Most of the previous works resort to using the Chamfer discrepancy or Earth Mover's distance, but those metrics are either ineffective in measuring the differences between point clouds or computationally expensive. In this paper, we conduct a systematic study with extensive experiments on distance metrics for 3D point clouds. From this study, we propose to use sliced Wasserstein distance and its variants for learning representations of 3D point clouds. In addition, we introduce a new algorithm to estimate sliced Wasserstein distance that guarantees that the estimated value is close enough to the true one. Experiments show that the sliced Wasserstein distance and its variants allow the neural network to learn a more efficient representation compared to the Chamfer discrepancy. We demonstrate the efficiency of the sliced Wasserstein metric and its variants on several tasks in 3D computer vision including training a point cloud autoencoder, generative modeling, transfer learning, and point cloud registration.

We propose a new approach to human clothing modeling based on point clouds. Within this approach, we learn a deep model that can predict point clouds of various outfits, for various human poses, and for various human body shapes. Notably, outfits of various types and topologies can be handled by the same model. Using the learned model, we can infer the geometry of new outfits from as little as a single image, and perform outfit retargeting to new bodies in new poses. We complement our geometric model with appearance modeling that uses the point cloud geometry as a geometric scaffolding and employs neural point-based graphics to capture outfit appearance from videos and to re-render the captured outfits. We validate both geometric modeling and appearance modeling aspects of the proposed approach against recently proposed methods and establish the viability of point-based clothing modeling.

Self-attention networks have revolutionized natural language processing and are making impressive strides in image analysis tasks such as image classification and object detection. Inspired by this success, we investigate the application of self-attention networks to 3D point cloud processing. We design self-attention layers for point clouds and use these to construct self-attention networks for tasks such as semantic scene segmentation, object part segmentation, and object classification. Our Point Transformer design improves upon prior work across domains and tasks. For example, on the challenging S3DIS dataset for large-scale semantic scene segmentation, the Point Transformer attains an mIoU of 70.4% on Area 5, outperforming the strongest prior model by 3.3 absolute percentage points and crossing the 70% mIoU threshold for the first time.

Despite the extensive usage of point clouds in 3D vision, relatively limited data are available for training deep neural networks. Although data augmentation is a standard approach to compensate for the scarcity of data, it has been less explored in the point cloud literature. In this paper, we propose a simple and effective augmentation method called PointWOLF for point cloud augmentation. The proposed method produces smoothly varying non-rigid deformations by locally weighted transformations centered at multiple anchor points. The smooth deformations allow diverse and realistic augmentations. Furthermore, in order to minimize the manual efforts to search the optimal hyperparameters for augmentation, we present AugTune, which generates augmented samples of desired difficulties producing targeted confidence scores. Our experiments show that our framework consistently improves the performance for both shape classification and part segmentation tasks. In particular, with PointNet++, PointWOLF achieves the state-of-the-art 89.7 accuracy on shape classification with the real-world ScanObjectNN dataset. The code is available at https://github.com/mlvlab/PointWOLF.

Point clouds captured in real-world applications are often incomplete due to the limited sensor resolution, single viewpoint, and occlusion. Therefore, recovering the complete point clouds from partial ones becomes an indispensable task in many practical applications. In this paper, we present a new method that reformulates point cloud completion as a set-to-set translation problem and design a new model, called PoinTr that adopts a transformer encoder-decoder architecture for point cloud completion. By representing the point cloud as a set of unordered groups of points with position embeddings, we convert the point cloud to a sequence of point proxies and employ the transformers for point cloud generation. To facilitate transformers to better leverage the inductive bias about 3D geometric structures of point clouds, we further devise a geometry-aware block that models the local geometric relationships explicitly. The migration of transformers enables our model to better learn structural knowledge and preserve detailed information for point cloud completion. Furthermore, we propose two more challenging benchmarks with more diverse incomplete point clouds that can better reflect the real-world scenarios to promote future research. Experimental results show that our method outperforms state-of-the-art methods by a large margin on both the new benchmarks and the existing ones.

Accurate camera pose estimation or global camera re-localization is a core component in Structure-from-Motion (SfM) and SLAM systems. Given pair-wise relative camera poses, pose-graph optimization (PGO) involves solving for an optimized set of globally-consistent absolute camera poses. In this work, we propose a novel PGO scheme fueled by graph neural networks (GNN), namely PoGO-Net, to conduct the absolute camera pose regression leveraging multiple rotation averaging (MRA). Specifically, PoGO-Net takes a noisy view-graph as the input, where the nodes and edges are designed to encode the geometric constraints and local graph consistency. Besides, we address the outlier edge removal by exploiting an implicit edge-dropping scheme where the noisy or corrupted edges are effectively filtered out with parameterized networks. Furthermore, we introduce a joint loss function embedding MRA formulation such that the robust inference is capable of achieving real-time performances even for large-scale scenes. Our proposed network is trained end-to-end on public benchmarks, outperforming state-of-the-art approaches in extensive experiments that demonstrate the efficiency and robustness of our proposed network.

Recently, DETR pioneered the solution of vision tasks with transformers, it directly translates the image feature map into the object detection result. Though effective, translating the full feature map can be costly due to redundant computation on some area like the background. In this work, we encapsulate the idea of reducing spatial redundancy into a novel poll and pool (PnP) sampling module, with which we build an end-to-end PnP-DETR architecture that adaptively allocates its computation spatially to be more efficient. Concretely, the PnP module abstracts the image feature map into fine foreground object feature vectors and a small number of coarse background contextual feature vectors. The transformer models information interaction within the fine-coarse feature space and translates the features into the detection result. Moreover, the PnP-augmented model can instantly achieve various desired trade-offs between performance and computation with a single model by varying the sampled feature length, without requiring to train multiple models as existing methods. Thus it offers greater flexibility for deployment in diverse scenarios with varying computation constraint. We further validate the generalizability of the PnP module on panoptic segmentation and the recent transformer-based image recognition model ViT and show consistent efficiency gain. We believe our method makes a step for efficient visual analysis with transformers, wherein spatial redundancy is commonly observed. Code and models will be available.

We introduce a method to render Neural Radiance Fields (NeRFs) in real time using PlenOctrees, an octree-based 3D representation which supports view-dependent effects. Our method can render 800x800 images at more than 150 FPS, which is over 3000 times faster than conventional NeRFs. We do so without sacrificing quality while preserving the ability of NeRFs to perform free-viewpoint rendering of scenes with arbitrary geometry and view-dependent effects. Real-time performance is achieved by pre-tabulating the NeRF into a PlenOctree. In order to preserve view-dependent effects such as specularities, we factorize the appearance via closed-form spherical basis functions. Specifically, we show that it is possible to train NeRFs to predict a spherical harmonic representation of radiance, removing the viewing direction as an input to the neural network. Furthermore, we show that PlenOctrees can be directly optimized to further minimize the reconstruction loss, which leads to equal or better quality compared to competing methods. Moreover, this octree optimization step can be used to reduce the training time, as we no longer need to wait for the NeRF training to converge fully. Our real-time neural rendering approach may potentially enable new applications such as 6-DOF industrial and product visualizations, as well as next generation AR/VR systems. PlenOctrees are amenable to in-browser rendering as well; please visit the project page for the interactive online demo, as well as video and code: https://alexyu.net/plenoctrees.

This paper presents a neural network built upon Transformers, namely PlaneTR, to simultaneously detect and reconstruct planes from a single image. Different from previous methods, PlaneTR jointly leverages the context information and the geometric structures in a sequence-to-sequence way to holistically detect plane instances in one forward pass. Specifically, we represent the geometric structures as line segments and conduct the network with three main components: (i) context and line segments encoders, (ii) a structure-guided plane decoder, (iii) a pixel-wise plane embedding decoder. Given an image and its detected line segments, PlaneTR generates the context and line segment sequences via two specially designed encoders and then feeds them into a Transformers-based decoder to directly predict a sequence of plane instances by simultaneously considering the context and global structure cues. Finally, the pixel-wise embeddings are computed to assign each pixel to one predicted plane instance which is nearest to it in embedding space. Comprehensive experiments demonstrate that PlaneTR achieves state-of-the-art performance on the ScanNet and NYUv2 datasets.

The paper studies planar surface reconstruction of indoor scenes from two views with unknown camera poses. While prior approaches have successfully created object-centric reconstructions of many scenes, they fail to exploit other structures, such as planes, which are typically the dominant components of indoor scenes. In this paper, we reconstruct planar surfaces from multiple views, while jointly estimating camera pose. Our experiments demonstrate that our method is able to advance the state of the art of reconstruction from sparse views, on challenging scenes from Matterport3D.

Recent advancements in differentiable rendering and 3D reasoning have driven exciting results in novel view synthesis from a single image. Despite realistic results, methods are limited to relatively small view change. In order to synthesize immersive scenes, models must also be able to extrapolate. We present an approach that fuses 3D reasoning with autoregressive modeling to outpaint large view changes in a 3D-consistent manner, which enables scene synthesis. We demonstrate considerable improvement in single-image large-angle view synthesis results compared to a variety of methods and possible variants across simulated and real datasets. In addition, we show increased 3D consistency compared to alternative accumulation methods.

Autoregressive models are a class of exact inference approaches with highly flexible functional forms, yielding state-of-the-art density estimates for natural images. Yet, the sequential ordering on the dimensions makes these models computationally expensive and limits their applicability to low-resolution imagery. In this work, we propose Pixel-Pyramids, a block-autoregressive approach employing a lossless pyramid decomposition with scale-specific representations to encode the joint distribution of image pixels. Crucially, it affords a sparser dependency structure compared to fully autoregressive approaches. Our PixelPyramids yield state-of-the-art results for density estimation on various image datasets, especially for high-resolution data. For CelebA-HQ 1024 x 1024, we observe that the density estimates (in terms of bits/dim) are improved to 44% of the baseline despite sampling speeds superior even to easily parallelizable flow-based models.

Finding local features that are repeatable across multiple views is a cornerstone of sparse 3D reconstruction. The classical image matching paradigm detects keypoints per-image once and for all, which can yield poorly-localized features and propagate large errors to the final geometry. In this paper, we refine two key steps of structure-from-motion by a direct alignment of low-level image information from multiple views: we first adjust the initial keypoint locations prior to any geometric estimation, and subsequently refine points and camera poses as a post-processing. This refinement is robust to large detection noise and appearance changes, as it optimizes a featuremetric error based on dense features predicted by a neural network. This significantly improves the accuracy of camera poses and scene geometry for a wide range of keypoint detectors, challenging viewing conditions, and off-the-shelf deep features. Our system easily scales to large image collections, enabling pixel-perfect crowd-sourced localization at scale. Our code is publicly available at https://github.com/cvg/pixel-perfect-sfm as an add-on to the popular SfM software COLMAP.

Recently, deep Convolutional Neural Networks (CNNs) can achieve human-level performance in edge detection with the rich and abstract edge representation capacities. However, the high performance of CNN based edge detection is achieved with a large pretrained CNN backbone, which is memory and energy consuming. In addition, it is surprising that the previous wisdom from the traditional edge detectors, such as Canny, Sobel, and LBP are rarely investigated in the rapid-developing deep learning era. To address these issues, we propose a simple, lightweight yet effective architecture named Pixel Difference Network (PiDiNet) for efficient edge detection. Extensive experiments on BSDS500, NYUD, and Multicue are provided to demonstrate its effectiveness, and its high training and inference efficiency. Surprisingly, when training from scratch with only the BSDS500 and VOC datasets, PiDiNet can surpass the recorded result of human perception (0.807 vs. 0.803 in ODS F-measure) on the BSDS500 dataset with 100 FPS and less than 1M parameters. A faster version of PiDiNet with less than 0.1M parameters can still achieve comparable performance among state of the arts with 200 FPS. Results on the NYUD and Multicue datasets show similar observations. The codes are available at https://github.com/zhuoinoulu/pidinet.

We present a novel semi-supervised semantic segmentation method which jointly achieves two desiderata of segmentation model regularities: the label-space consistency property between image augmentations and the feature-space contrastive property among different pixels. We leverage the pixel-level L2 loss and the pixel contrastive loss for the two purposes respectively. To address the computational efficiency issue and the false negative noise issue involved in the pixel contrastive loss, we further introduce and investigate several negative sampling techniques. Extensive experiments demonstrate the state-of-the-art performance of our method (PC2Seg) with the DeepLab-v3+ architecture, in several challenging semi-supervised settings derived from the VOC, Cityscapes, and COCO datasets.

Recently proposed neural architecture search (NAS) methods co-train billions of architectures in a supernet and estimate their potential accuracy using the network weights detached from the supernet. However, the ranking correlation between the architectures' predicted accuracy and their actual capability is incorrect, which causes the existing NAS methods' dilemma. We attribute this ranking correlation problem to the supernet training consistency shift, including feature shift and parameter shift. Feature shift is identified as dynamic input distributions of a hidden layer due to random path sampling. The input distribution dynamic affects the loss descent and finally affects architecture ranking. Parameter shift is identified as contradictory parameter updates for a shared layer lay in different paths in different training steps. The rapidly-changing parameter could not preserve architecture ranking. We address these two shifts simultaneously using a nontrivial supernet-\Pi model, called \Pi-NAS. Specifically, we employ a supernet-\Pi model that contains cross-path learning to reduce the feature consistency shift between different paths. Meanwhile, we adopt a novel nontrivial mean teacher containing negative samples to overcome parameter shift and model collision. Furthermore, our \Pi-NAS runs in an unsupervised manner, which can search for more transferable architectures. Extensive experiments on ImageNet and a wide range of downstream tasks (e.g., COCO 2017, ADE20K, and Cityscapes) demonstrate the effectiveness and universality of our \Pi-NAS compared to supervised NAS. See Codes: https://github.com/Ernie1/Pi-NAS.

This paper introduces a new method of data-driven microscope design for virtual fluorescence microscopy. We use a deep neural network (DNN) to effectively design optical patterns for specimen illumination that substantially improve upon the ability to infer fluorescence image information from unstained microscope images. To achieve this design, we include an illumination model within the DNN's first layers that is jointly optimized during network training. We validated our method on two different experimental setups, with different magnifications and sample types, to show a consistent improvement in performance as compared to conventional microscope imaging methods. Additionally, to understand the importance of learned illumination on the inference task, we varied the number of illumination patterns being optimized (and thus the number of unique images captured) and analyzed how the structure of the patterns changed as their number increased. This work demonstrates the power of programmable optical elements at enabling better machine learning algorithm performance and at providing physical insight into next generation of machine-controlled imaging systems.

Human motion synthesis is an important problem for applications in graphics and gaming, and even in simulation environments for robotics. Existing methods require accurate motion capture data for training, which is costly to obtain. Instead, we propose a framework for training generative models of physically plausible human motion directly from monocular RGB videos, which are much more widely available. At the core of our method is a novel optimization formulation that aims to correct imperfect image-based pose estimations by enforcing physics constraints and reasons about contacts in a differentiable way. This optimization yields corrected 3D poses and motions, as well as their corresponding contact forces. Results show that our physically-correct motions significantly outperform prior work on pose estimation. We then train a generative model to synthesize both future motion and contact forces. We demonstrate both qualitatively and quantitatively significantly improved motion synthesis quality and physical plausibility achieved by our method on the large scale Human3.6m dataset as compared to prior learning-based kinematic and physics-based methods. By learning directly from video, our method paves the way for large-scale, realistic and diverse motion synthesis not previously possible.