ICCV 2021 论文列表

We consider class incremental learning (CIL) problem, in which a learning agent continuously learns new classes from incrementally arriving training data batches and aims to predict well on all the classes learned so far. The main challenge of the problem is the catastrophic forgetting, and for the exemplar-memory based CIL methods, it is generally known that the forgetting is commonly caused by the classification score bias that is injected due to the data imbalance between the new classes and the old classes (in the exemplar-memory). While several methods have been proposed to correct such score bias by some additional post-processing, e.g., score re-scaling or balanced fine-tuning, no systematic analysis on the root cause of such bias has been done. To that end, we analyze that computing the softmax probabilities by combining the output scores for all old and new classes could be the main cause of the bias. Then, we propose a new CIL method, dubbed as Separated Softmax for Incremental Learning (SS-IL), that consists of separated softmax (SS) output layer combined with task-wise knowledge distillation (TKD) to resolve such bias. Throughout our extensive experimental results on several large-scale CIL benchmark datasets, we show our SS-IL achieves strong state-of-the-art accuracy through attaining much more balanced prediction scores across old and new classes, without any additional post-processing.

For unsupervised image-to-image translation, we propose a discriminator architecture which focuses on the statistical features instead of individual patches. The network is stabilized by distribution matching of key statistical features at multiple scales. Unlike the existing methods which impose more and more constraints on the generator, our method facilitates the shape deformation and enhances the fine details with a greatly simplified framework. We show that the proposed method outperforms the existing state-of-the-art models in various challenging applications including selfie-to-anime, male-to-female and glasses removal.

In autonomous driving, a LiDAR-based object detector should perform reliably at different geographic locations and under various weather conditions. While recent 3D detection research focuses on improving performance within a single domain, our study reveals that the performance of modern detectors can drop drastically cross-domain. In this paper, we investigate unsupervised domain adaptation (UDA) for LiDAR-based 3D object detection. On the Waymo Domain Adaptation dataset, we identify the deteriorating point cloud quality as the root cause of the performance drop. To address this issue, we present Semantic Point Generation (SPG), a general approach to enhance the reliability of LiDAR detectors against domain shifts. Specifically, SPG generates semantic points at the predicted foreground regions and faithfully recovers missing parts of the foreground objects, which are caused by phenomena such as occlusions, low reflectance, or weather interference. By merging the semantic points with the original points, we obtain an augmented point cloud, which can be directly consumed by modern LiDAR-based detectors. To validate the wide applicability of SPG, we experiment with two representative detectors, PointPillars and PV-RCNN. On the UDA task, SPG significantly improves both detectors across all object categories of interest and at all difficulty levels. SPG can also benefit object detection in the original domain. On the Waymo Open Dataset and KITTI, SPG improves 3D detection results of these two methods across all categories. Combined with PV-RCNN, SPG achieves state-of-the-art 3D detection results on KITTI.

Due to the lack of camera parameter information for in-the-wild images, existing 3D human pose and shape (HPS) estimation methods make several simplifying assumptions: weak-perspective projection, large constant focal length, and zero camera rotation. These assumptions often do not hold and we show, quantitatively and qualitatively, that they cause errors in the reconstructed 3D shape and pose. To address this, we introduce SPEC, the first in-the-wild 3D HPS method that estimates the perspective camera from a single image and employs this to reconstruct 3D human bodies more accurately. First, we train a neural network to estimate the field of view, camera pitch, and roll given an input image. We employ novel losses that improve the calibration accuracy over previous work. We then train a novel network that concatenates the camera calibration to the image features and uses these together to regress 3D body shape and pose. SPEC is more accurate than the prior art on the standard benchmark (3DPW) as well as two new datasets with more challenging camera views and varying focal lengths. Specifically, we create a new photorealistic synthetic dataset (SPEC-SYN) with ground truth 3D bodies and a novel in-the-wild dataset (SPEC-MTP) with calibration and high-quality reference bodies. Code and datasets are available for research purposes at https://spec.is.tue.mpg.de/.

Most recent transformer-based models show impressive performance on vision tasks, even better than Convolution Neural Networks (CNN). In this work, we present a novel, flexible, and effective transformer-based model for high-quality instance segmentation. The proposed method, Segmenting Objects with TRansformers (SOTR), simplifies the segmentation pipeline, building on an alternative CNN backbone appended with two parallel subtasks: (1) predicting per-instance category via transformer and (2) dynamically generating segmentation mask with the multi-level upsampling module. SOTR can effectively extract lower-level feature representations and capture long-range context dependencies by Feature Pyramid Network (FPN) and twin transformer, respectively. Meanwhile, compared with the original transformer, the proposed twin transformer is timeand resource-efficient since only a row and a column attention are involved to encode pixels. Moreover, SOTR is easy to be incorporated with various CNN backbones and transformer model variants to make considerable improvements for the segmentation accuracy and training convergence. Extensive experiments show that our SOTR performs well on the MS COCO dataset and surpasses state-of-the-art instance segmentation approaches. We hope our simple but strong framework could serve as a preferment baseline for instance-level recognition. Our code is available at https://github.com/easton-cau/SOTR.

Marker-based optical motion capture (mocap) is the "gold standard" method for acquiring accurate 3D human motion in computer vision, medicine, and graphics. The raw output of these systems are noisy and incomplete 3D points or short tracklets of points. To be useful, one must associate these points with corresponding markers on the captured subject; i.e. "labelling". Given these labels, one can then "solve" for the 3D skeleton or body surface mesh. Commercial auto-labeling tools require a specific calibration procedure at capture time, which is not possible for archival data. Here we train a novel neural network called SOMA, which takes raw mocap point clouds with varying numbers of points, labels them at scale without any calibration data, independent of the capture technology, and requiring only minimal human intervention. Our key insight is that, while labeling point clouds is highly ambiguous, the 3D body provides strong constraints on the solution that can be exploited by a learning-based method. To enable learning, we generate massive training sets of simulated noisy and ground truth mocap markers animated by 3D bodies from AMASS. SOMA exploits an architecture with stacked self-attention elements to learn the spatial structure of the 3D body and an optimal transport layer to constrain the assignment (labeling) problem while rejecting outliers. We extensively evaluate SOMA both quantitatively and qualitatively. SOMA is more accurate and robust than existing state of the art research methods and can be applied where commercial systems cannot. We automatically label over 8 hours of archival mocap data across 4 different datasets captured using various technologies and output SMPL-X body models. The model and data is released for research purposes at https://soma.is.tue.mpg.de/.

Directly regressing all 6 degrees-of-freedom (6DoF) for the object pose (i.e. the 3D rotation and translation) in a cluttered environment from a single RGB image is a challenging problem. While end-to-end methods have recently demonstrated promising results at high efficiency, they are still inferior when compared with elaborate PnP/RANSAC-based approaches in terms of pose accuracy. In this work, we address this shortcoming by means of a novel reason-ing about self-occlusion, in order to establish a two-layer representation for 3D objects which considerably enhances the accuracy of end-to-end 6D pose estimation. Our frame-work, named SO-Pose, takes a single RGB image as input and respectively generates 2D-3D correspondences as well as self-occlusion information harnessing a shared encoder and two separate decoders. Both outputs are then fused to directly regress the 6DoF pose parameters. Incorporating cross-layer consistencies that align correspondences, self-occlusion, and 6D pose, we can further improve accuracy and robustness, surpassing or rivaling all other state-of-the-art approaches on various challenging datasets.

Neural implicit surface representations have emerged as a promising paradigm to capture 3D shapes in a continuous and resolution-independent manner. However, adapting them to articulated shapes is non-trivial. Existing approaches learn a backward warp field that maps deformed to canonical points. However, this is problematic since the backward warp field is pose dependent and thus requires large amounts of data to learn. To address this, we introduce SNARF, which combines the advantages of linear blend skinning (LBS) for polygonal meshes with those of neural implicit surfaces by learning a forward deformation field without direct supervision. This deformation field is defined in canonical, pose-independent, space, enabling generalization to unseen poses. Learning the deformation field from posed meshes alone is challenging since the correspondences of deformed points are defined implicitly and may not be unique under changes of topology. We propose a forward skinning model that finds all canonical correspondences of any deformed point using iterative root finding. We derive analytical gradients via implicit differentiation, enabling end-to-end training from 3D meshes with bone transformations. Compared to state-of-the-art neural implicit representations, our approach generalizes better to unseen poses while preserving accuracy. We demonstrate our method in challenging scenarios on (clothed) 3D humans in diverse and unseen poses.

Recently, several frameworks for self-supervised learning of 3D scene flow on point clouds have emerged. Scene flow inherently separates every scene into multiple moving agents and a large class of points following a single rigid sensor motion. However, existing methods do not leverage this property of the data in their self-supervised training routines which could improve and stabilize flow predictions. Based on the discrepancy between a robust rigid ego-motion estimate and a raw flow prediction, we generate a self-supervised motion segmentation signal. The predicted motion segmentation, in turn, is used by our algorithm to attend to stationary points for aggregation of motion information in static parts of the scene. We learn our model end-to-end by backpropagating gradients through Kabsch's algorithm and demonstrate that this leads to accurate ego-motion which in turn improves the scene flow estimate. Using our method, we show state-of-the-art results across multiple scene flow metrics for different real-world datasets, showcasing the robustness and generalizability of this approach. We further analyze the performance gain when performing joint motion segmentation and scene flow in an ablation study. We also present a novel network architecture for 3D LiDAR scene flow which is capable of handling an order of magnitude more points during training than previously possible.

Single image 3D photography enables viewers to view a still image from novel viewpoints. Recent approaches combine monocular depth networks with inpainting networks to achieve compelling results. A drawback of these techniques is the use of hard depth layering, making them unable to model intricate appearance details such as thin hair-like structures. We present SLIDE, a modular and unified system for single image 3D photography that uses a simple yet effective soft layering strategy to better preserve appearance details in novel views. In addition, we propose a novel depth-aware training strategy for our inpainting module, better suited for the 3D photography task. The resulting SLIDE approach is modular, enabling the use of other components such as segmentation and matting for improved layering. At the same time, SLIDE uses an efficient layered depth formulation that only requires a single forward pass through the component networks to produce high quality 3D photos. Extensive experimental analysis on three view-synthesis datasets, in combination with user studies on in-the-wild image collections, demonstrate superior performance of our technique in comparison to existing strong baselines while being conceptually much simpler. Project page: https://varunjampani.github.io/slide

Motion is an important cue for video prediction and often utilized by separating video content into static and dynamic components. Most of the previous work utilizing motion is deterministic but there are stochastic methods that can model the inherent uncertainty of the future. Existing stochastic models either do not reason about motion explicitly or make limiting assumptions about the static part. In this paper, we reason about appearance and motion in the video stochastically by predicting the future based on the motion history. Explicit reasoning about motion without history already reaches the performance of current stochastic models. The motion history further improves the results by allowing to predict consistent dynamics several frames into the future. Our model performs comparably to the state-of-the-art models on the generic video prediction datasets, however, significantly outperforms them on two challenging real-world autonomous driving datasets with complex motion and dynamic background.

By estimating 3D shape and instances from a single view, we can capture information about the environment quickly, without the need for comprehensive scanning and multi-view fusion. Solving this task for composite scenes (such as object stacks) is challenging: occluded areas are not only ambiguous in shape but also in instance segmentation; multiple decompositions could be valid. We observe that physics constrains decomposition as well as shape in occluded regions and hypothesise that a latent space learned from scenes built under physics simulation can serve as a prior to better predict shape and instances in occluded regions. To this end we propose SIMstack, a depth-conditioned Variational Auto-Encoder (VAE), trained on a dataset of objects stacked under physics simulation. We formulate instance segmentation as a center voting task which allows for class-agnostic detection and doesn't require setting the maximum number of objects in the scene. At test time, our model can generate 3D shape and instance segmentation from a single depth view, probabilistically sampling proposals for the occluded region from the learned latent space. We argue that this method has practical applications in providing robots some of the ability humans have to make rapid intuitive inferences of partially observed scenes. We demonstrate an application for precise (non-disruptive) object grasping of unknown objects from a single depth view.

We present a novel neural representation for light field content that enables compact storage and easy local reconstruction with high fidelity. We use a fully-connected neural network to learn the mapping function between each light field pixel's coordinates and its corresponding color values. However, neural networks that simply take in raw coordinates are unable to accurately learn data containing fine details. We present an input transformation strategy based on the Gegenbauer polynomials which previously showed theoretical advantages over the Fourier basis. We conduct experiments that show our Gegenbauer-based design combined with sinusoidal activation functions leads to a better light field reconstruction quality than a variety of network designs, including those with Fourier-inspired techniques introduced by prior works. Moreover, our SInusoidal Gegenbauer NETwork, or SIGNET, can represent light field scenes more compactly than the state-of-the-art compression methods while maintaining a comparable reconstruction quality. SIGNET also innately allows random access to encoded light field pixels due to its functional design. Furthermore, we demonstrate that SIGNET facilitates super-resolution along the spatial, angular, and temporal dimensions of a light field without any additional training.

Unlike conventional zero-shot classification, zero-shot semantic segmentation predicts a class label at the pixel level instead of the image level. When solving zero-shot semantic segmentation problems, the need for pixel-level prediction with surrounding context motivates us to incorporate spatial information using positional encoding. We improve standard positional encoding by introducing the concept of Relative Positional Encoding, which integrates spatial information at the feature level and can handle arbitrary image sizes. Furthermore, while self-training is widely used in zero-shot semantic segmentation to generate pseudo-labels, we propose a new knowledge-distillation-inspired self-training strategy, namely Annealed Self-Training, which can automatically assign different importance to pseudo-labels to improve performance. We systematically study the proposed Relative Positional Encoding and Annealed Self-Training in a comprehensive experimental evaluation, and our empirical results confirm the effectiveness of our method on three benchmark datasets.

Category-level 6D object pose estimation aims to predict the position and orientation for unseen objects, which plays a pillar role in many scenarios such as robotics and augmented reality. The significant intra-class variation is the bottleneck challenge in this task yet remains unsolved so far. In this paper, we take advantage of category prior to overcome this problem by innovating a structure-guided prior adaptation scheme to accurately estimate 6D pose for individual objects. Different from existing prior-based methods, given one object and its corresponding category prior, we propose to leverage their structure similarity to dynamically adapt the prior to the observed object. The prior adaptation intrinsically associates the adopted prior with different objects, from which we can accurately reconstruct the 3D canonical model of the specific object for pose estimation. To further enhance the structure characteristic of objects, we extract low-rank structure points from the dense object point cloud, therefore more efficiently incorporating sparse structural information during prior adaptation. Extensive experiments on CAMERA25 and REAL275 benchmarks demonstrate significant performance improvement. Project homepage: https://www.cse.cuhk.edu.hk/ kaichen/projects/sgpa/sgpa.html.

Recently, various works that attempted to introduce rotation invariance to point cloud analysis have devised point-pair features, such as angles and distances. In these methods, however, the point-pair is only comprised of the center point and its adjacent points in a vicinity, which may bring information loss to the local feature representation. In this paper, we instead connect each point densely with all other points in a local neighborhood to compose the point-pairs. Specifically, we present a simple but effective local feature representation, called sorted Gram matrix(SGM), which is not only invariant to arbitrary rotations, but also models the pair-wise relationship of all the points in a neighborhood. In more detail, we utilize vector inner product to model distance- and angle-information between two points, and in a local patch it naturally forms a Gram matrix. In order to guarantee permutation invariance, we sort the correlation value in Gram matrix for each point, therefore this geometric feature names sorted Gram matrix. Furthermore, we mathematically prove that the Gram matrix is rotation-invariant and sufficient to model the inherent structure of a point cloud patch. We then use SGM as features in convolution, which can be readily integrated as a drop-in module into any point-based networks. Finally, we evaluated the proposed method on two widely used datasets, and it outperforms previous state-of-the-arts on both shape classification and part segmentation tasks by a large margin.

Many existing approaches for unsupervised domain adaptation (UDA) focus on adapting under only data distribution shift and offer limited success under additional cross-domain label distribution shift. Recent work based on self-training using target pseudolabels has shown promise, but on challenging shifts pseudolabels may be highly unreliable and using them for self-training may lead to error accumulation and domain misalignment. We propose Selective Entropy Optimization via Committee Consistency (SENTRY), a UDA algorithm that judges the reliability of a target instance based on its predictive consistency under a committee of random image transformations. Our algorithm then selectively minimizes predictive entropy to increase confidence on highly consistent target instances, while maximizing predictive entropy to reduce confidence on highly inconsistent ones. In combination with pseudolabel-based approximate target class balancing, our approach leads to significant improvements over the state-of-the-art on 27/31 domain shifts from standard UDA benchmarks as well as benchmarks designed to stress-test adaptation under label distribution shift.

Explainable artificial intelligence has been gaining attention in the past few years. However, most existing methods are based on gradients or intermediate features, which are not directly involved in the decision-making process of the classifier. In this paper, we propose a slot attention-based classifier called SCOUTER for transparent yet accurate classification. Two major differences from other attention-based methods include: (a) SCOUTER's explanation is involved in the final confidence for each category, offering more intuitive interpretation, and (b) all the categories have their corresponding positive or negative explanation, which tells "why the image is of a certain category" or "why the image is not of a certain category." We design a new loss tailored for SCOUTER that controls the model's behavior to switch between positive and negative explanations, as well as the size of explanatory regions. Experimental results show that SCOUTER can give better visual explanations in terms of various metrics while keeping good accuracy on small and medium-sized datasets.

3D visual grounding aims at grounding a natural language description about a 3D scene, usually represented in the form of 3D point clouds, to the targeted object region. Point clouds are sparse, noisy, and contain limited semantic information compared with 2D images. These inherent limitations make the 3D visual grounding problem more challenging. In this study, we propose 2D Semantics Assisted Training (SAT) that utilizes 2D image semantics in the training stage to ease point-cloud-language joint representation learning and assist 3D visual grounding. The main idea is to learn auxiliary alignments between rich, clean 2D object representations and the corresponding objects or mentioned entities in 3D scenes. SAT takes 2D object semantics, i.e., object label, image feature, and 2D geometric feature, as the extra input in training but does not require such inputs during inference. By effectively utilizing 2D semantics in training, our approach boosts the accuracy on the Nr3D dataset from 37.7% to 49.2%, which significantly surpasses the non-SAT baseline with the identical network architecture and inference input. Our approach outperforms the state of the art by large margins on multiple 3D visual grounding datasets, i.e., +10.4% absolute accuracy on Nr3D, +9.9% on Sr3D, and +5.6% on ScanRef.