ICCV 2021 论文列表

We present ALADIN (All Layer AdaIN); a novel architecture for searching images based on the similarity of their artistic style. Representation learning is critical to visual search, where distance in the learned search embedding reflects image similarity. Learning an embedding that discriminates fine-grained variations in style is hard, due to the difficulty of defining and labelling style. ALADIN takes a weakly supervised approach to learning a representation for fine-grained style similarity of digital artworks, leveraging BAM-FG, a novel large-scale dataset of user generated content groupings gathered from the web. ALADIN sets a new state of the art accuracy for style-based visual search over both coarse labelled style data (BAM) and BAM-FG; a new 2.62 million image dataset of 310,000 fine-grained style groupings also contributed by this work.

Recently, some approaches are proposed to harness deep convolutional networks to facilitate superpixel segmentation. The common practice is to first evenly divide the image into a pre-defined number of grids and then learn to associate each pixel with its surrounding grids. However, simply applying a series of convolution operations with limited receptive fields can only implicitly perceive the relations between the pixel and its surrounding grids. Consequently, existing methods often fail to provide an effective context when inferring the association map. To remedy this issue, we propose a novel Association Implantation (AI) module to enable the network to explicitly capture the relations between the pixel and its surrounding grids. The proposed AI module directly implants the features of grid cells to the surrounding of its corresponding central pixel, and conducts convolution on the padded window to adaptively transfer knowledge between them. With such an implantation operation, the network could explicitly harvest the pixel-grid level context, which is more in line with the target of superpixel segmentation comparing to the pixel-wise relation. Furthermore, to pursue better boundary precision, we design a boundary-perceiving loss to help the network discriminate the pixels around boundaries in hidden feature level, which could benefit the subsequent inferring modules to accurately identify more boundary pixels. Extensive experiments on BSDS500 and NYUv2 datasets show that our method could not only achieve state-of-the-art performance but maintain satisfactory inference efficiency. Code and pre-trained model are available at https://github.com/wangyxxjtu/AINet-ICCV2021.

We present AIST++, a new multi-modal dataset of 3D dance motion and music, along with FACT, a Full-Attention Cross-modal Transformer network for generating 3D dance motion conditioned on music. The proposed AIST++ dataset contains 1.1M frames of 3D dance motion in 1408 sequences, covering 10 dance genres with multi-view videos with known camera poses---the largest dataset of this kind to our knowledge. We show that naively applying sequence models such as transformers to this dataset for the task of music conditioned 3D motion generation does not produce satisfactory 3D motion that is well correlated with the input music. We overcome these shortcomings by introducing key changes in its architecture design and supervision: FACT model involves a deep cross-modal transformer block with full-attention that is trained to predict N future motions. We empirically show that these changes are key factors in generating long sequences of realistic dance motion that are well-attuned to the input music. We conduct extensive experiments on AIST++ with user studies, where our method outperforms recent state-of-the-art methods both qualitatively and quantitatively. The code and the dataset can be found at: https://google.github.io/aichoreographer.

While deep neural networks have shown impressive performance in many tasks, they are fragile to carefully designed adversarial attacks. We propose a novel adversarial training-based model by Attention Guided Knowledge Distillation and Bi-directional Metric Learning (AGKD-BML). The attention knowledge is obtained from a weight-fixed model trained on a clean dataset, referred to as a teacher model, and transferred to a model that is under training on adversarial examples (AEs), referred to as a student model. In this way, the student model is able to focus on the correct region, as well as correcting the intermediate features corrupted by AEs to eventually improve the model accuracy. Moreover, to efficiently regularize the representation in feature space, we propose a bidirectional metric learning. Specifically, given a clean image, it is first attacked to its most confusing class to get the forward AE. A clean image in the most confusing class is then randomly picked and attacked back to the original class to get the backward AE. A triplet loss is then used to shorten the representation distance between original image and its AE, while enlarge that between the forward and backward AEs. We conduct extensive adversarial robustness experiments on two widely used datasets with different attacks. Our proposed AGKD-BML model consistently outperforms the state-of-the-art approaches. The code of AGKD-BML will be available at: https://github.com/hongw579/AGKD-BML.

Visual storytelling and story comprehension are uniquely human skills that play a central role in how we learn about and experience the world. Despite remarkable progress in recent years in synthesis of visual and textual content in isolation and learning effective joint visual-linguistic representations, existing systems still operate only at a superficial, factual level. With the goal of developing systems that are able to comprehend rich human-generated narratives, and co-create new stories, we introduce AESOP: a new dataset that captures the creative process associated with visual storytelling. Visual panels are composed of clip-art objects with specific attributes enabling a broad range of creative expression. Using AESOP, we propose foundational storytelling tasks that are generative variants of story cloze tests, to better measure the creative and causal reasoning ability required for visual storytelling. We further develop a generalized story completion framework that models stories as the co-evolution of visual and textual concepts. We benchmark the proposed approach with human baselines and evaluate using comprehensive qualitative and quantitative metrics. Our results highlight key insights related to the dataset, modelling and evaluation of visual storytelling for future research in this promising field of study.

The recent progress of CNN has dramatically improved face alignment performance. However, few works have paid attention to the error-bias with respect to error distribution of facial landmarks. In this paper, we investigate the error-bias issue in face alignment, where the distributions of landmark errors tend to spread along the tangent line to landmark curves. This error-bias is not trivial since it is closely connected to the ambiguous landmark labeling task. Inspired by this observation, we seek a way to leverage the error-bias property for better convergence of CNN model. To this end, we propose anisotropic direction loss (ADL) and anisotropic attention module (AAM) for coordinate and heatmap regression, respectively. ADL imposes strong binding force in normal direction for each landmark point on facial boundaries. On the other hand, AAM is an attention module which can get anisotropic attention mask focusing on the region of point and its local edge connected by adjacent points, it has a stronger response in tangent than in normal, which means relaxed constraints in the tangent. These two methods work in a complementary manner to learn both facial structures and texture details. Finally, we integrate them into an optimized end-to-end training pipeline named ADNet. Our ADNet achieves state-of-the-art results on 300W, WFLW and COFW datasets, which demonstrates the effectiveness and robustness.

Generating high-fidelity talking head video by fitting with the input audio sequence is a challenging problem that receives considerable attentions recently. In this paper, we address this problem with the aid of neural scene representation networks. Our method is completely different from existing methods that rely on intermediate representations like 2D landmarks or 3D face models to bridge the gap between audio input and video output. Specifically, the feature of input audio signal is directly fed into a conditional implicit function to generate a dynamic neural radiance field, from which a high-fidelity talking-head video corresponding to the audio signal is synthesized using volume rendering. Another advantage of our framework is that not only the head (with hair) region is synthesized as previous methods did, but also the upper body is generated via two individual neural radiance fields. Experimental results demonstrate that our novel framework can (1) produce high-fidelity and natural results, and (2) support free adjustment of audio signals, viewing directions, and background images. Code is available at https://github.com/YudongGuo/AD-NeRF.

One-stage long-tailed recognition methods improve the overall performance in a "seesaw" manner, i.e., either sacrifice the head's accuracy for better tail classification or elevate the head's accuracy even higher but ignore the tail. Existing algorithms bypass such trade-off by a multi-stage training process: pre-training on imbalanced set and fine-tuning on balanced set. Though achieving promising performance, not only are they sensitive to the generalizability of the pre-trained model, but also not easily integrated into other computer vision tasks like detection and segmentation, where pre-training of classifier solely is not applicable. In this paper, we propose a one-stage long-tailed recognition scheme, ally complementary experts (ACE), where the expert is the most knowledgeable specialist in a sub-set that dominates its training, and is complementary to other experts in the less-seen categories without disturbed by what it has never seen. We design a distribution-adaptive optimizer to adjust the learning pace of each expert to avoid over-fitting. Without special bells and whistles, the vanilla ACE outperforms the current one-stage SOTA method by 3 10% on CIFAR10-LT, CIFAR100-LT, ImageNet-LT and iNaturalist datasets. It is also shown to be the first one to break the "seesaw" trade-off by improving the accuracy of the majority and minority categories simultaneously in only one stage.

Level 5 autonomy for self-driving cars requires a robust visual perception system that can parse input images under any visual condition. However, existing semantic segmentation datasets are either dominated by images captured under normal conditions or are small in scale. To address this, we introduce ACDC, the Adverse Conditions Dataset with Correspondences for training and testing semantic segmentation methods on adverse visual conditions. ACDC consists of a large set of 4006 images which are equally distributed between four common adverse conditions: fog, nighttime, rain, and snow. Each adverse-condition image comes with a high-quality fine pixel-level semantic annotation, a corresponding image of the same scene taken under normal conditions, and a binary mask that distinguishes between intra-image regions of clear and uncertain semantic content. Thus, ACDC supports both standard semantic segmentation and the newly introduced uncertainty-aware semantic segmentation. A detailed empirical study demonstrates the challenges that the adverse domains of ACDC pose to state-of-the-art supervised and unsupervised approaches and indicates the value of our dataset in steering future progress in the field. Our dataset and benchmark are publicly available.

The natural association between visual observations and their corresponding sound provides powerful self-supervisory signals for learning video representations, which makes the ever-growing amount of online videos an attractive source of training data. However, large portions of online videos contain irrelevant audio-visual signals because of edited/overdubbed audio, and models trained on such uncurated videos have shown to learn suboptimal representations. Therefore, existing self-supervised approaches rely on datasets with predetermined taxonomies of semantic concepts, where there is a high chance of audio-visual correspondence. Unfortunately, constructing such datasets require labor intensive manual annotation and/or verification, which severely limits the utility of online videos for large-scale learning. In this work, we present an automatic dataset curation approach based on subset optimization where the objective is to maximize the mutual information between audio and visual channels in videos. We demonstrate that our approach finds videos with high audio-visual correspondence and show that self-supervised models trained on our data achieve competitive performances compared to models trained on existing manually curated datasets. The most significant benefit of our approach is scalability: We release ACAV100M that contains 100 million videos with high audio-visual correspondence, ideal for self-supervised video representation learning.

In this paper, we present a novel recurrent multi-view stereo network based on long short-term memory (LSTM) with adaptive aggregation, namely AA-RMVSNet. We firstly introduce an intra-view aggregation module to adaptively extract image features by using context-aware convolution and multi-scale aggregation, which efficiently improves the performance on challenging regions, such as thin objects and large low-textured surfaces. To overcome the difficulty of varying occlusion in complex scenes, we propose an inter-view cost volume aggregation module for adaptive pixel-wise view aggregation, which is able to preserve better-matched pairs among all views. The two proposed adaptive aggregation modules are lightweight, effective and complementary regarding improving the accuracy and completeness of 3D reconstruction. Instead of conventional 3D CNNs, we utilize a hybrid network with recurrent structure for cost volume regularization, which allows high-resolution reconstruction and finer hypothetical plane sweep. The proposed network is trained end-to-end and achieves excellent performance on various datasets. It ranks 1st among all submissions on Tanks and Temples benchmark and achieves competitive results on DTU dataset, which exhibits strong generalizability and robustness. Implementation of our method is available at https://github.com/QT-Zhu/AA-RMVSNet.

Recent work has made significant progress on using implicit functions, as a continuous representation for 3D rigid object shape reconstruction. However, much less effort has been devoted to modeling general articulated objects. Compared to rigid objects, articulated objects have higher degrees of freedom, which makes it hard to generalize to unseen shapes. To deal with the large shape variance, we introduce Articulated Signed Distance Functions (A-SDF) to represent articulated shapes with a disentangled latent space, where we have separate codes for encoding shape and articulation. With this disentangled continuous representation, we demonstrate that we can control the articulation input and animate unseen instances with unseen joint angles. Furthermore, we propose a Test-Time Adaptation inference algorithm to adjust our model during inference. We demonstrate our model generalize well to out-of-distribution and unseen data, e.g., partial point clouds and real-world depth images.

This paper addresses weakly supervised amodal instance segmentation, where the goal is to segment both visible and occluded (amodal) object parts, while training provides only ground-truth visible (modal) segmentations. Following prior work, we use data manipulation to generate occlusions in training images and thus train a segmenter to predict amodal segmentations of the manipulated data. The resulting predictions on training images are taken as the pseudo-ground truth for the standard training of Mask-RCNN, which we use for amodal instance segmentation of test images. For generating the pseudo-ground truth, we specify a new Amodal Segmenter based on Boundary Uncertainty estimation (ASBU) and make two contributions. First, while prior work uses the occluder's mask, our ASBU uses the occlusion boundary as input. Second, ASBU estimates an uncertainty map of the prediction. The estimated uncertainty regularizes learning such that lower segmentation loss is incurred on regions with high uncertainty. ASBU achieves significant performance improvement relative to the state of the art on the COCOA and KINS datasets in three tasks: amodal instance segmentation, amodal completion, and ordering recovery.

In this paper, we study the problem of Novel Class Discovery (NCD). NCD aims at inferring novel object categories in an unlabeled set by leveraging from prior knowledge of a labeled set containing different, but related classes. Existing approaches tackle this problem by considering multiple objective functions, usually involving specialized loss terms for the labeled and the unlabeled samples respectively, and often requiring auxiliary regularization terms. In this paper, we depart from this traditional scheme and introduce a UNified Objective function (UNO) for discovering novel classes, with the explicit purpose of favoring synergy between supervised and unsupervised learning. Using a multi-view self-labeling strategy, we generate pseudo-labels that can be treated homogeneously with ground truth labels. This leads to a single classification objective operating on both known and unknown classes. Despite its simplicity, UNO outperforms the state of the art by a significant margin on several benchmarks (approximately +10% on CIFAR-100 and +8% on ImageNet). Our source code will be publicly available. The project page is available at: https://ncd-uno.github.io.

We present a unified and flexible framework to address the generalized problem of 3D motion synthesis that covers the tasks of motion prediction, completion, interpolation, and spatial-temporal recovery. Since these tasks have different input constraints and various fidelity and diversity requirements, most existing approaches only cater to a specific task or use different architectures to address various tasks. Here we propose a unified framework based on Conditional Variational Auto-Encoder (CVAE), where we treat any arbitrary input as a masked motion series. Notably, by considering this problem as a conditional generation process, we estimate a parametric distribution of the missing regions based on the input conditions, from which to sample and synthesize the full motion series. To further allow the flexibility of manipulating the motion style of the generated series, we design an Action-Adaptive Modulation (AAM) to propagate the given semantic guidance through the whole sequence. We also introduce a cross-attention mechanism to exploit distant relations among decoder and encoder features for better realism and global consistency. We conducted extensive experiments on Human 3.6M and CMU-Mocap. The results show that our method produces coherent and realistic results for various motion synthesis tasks, with the synthesized motions distinctly adapted by the given action labels.

Mainstream state-of-the-art domain generalization algorithms tend to prioritize the assumption on semantic invariance across domains. Meanwhile, the inherent intra-domain style invariance is usually underappreciated and put on the shelf. In this paper, we reveal that leveraging intra-domain style invariance is also of pivotal importance in improving the efficiency of domain generalization. We verify that it is critical for the network to be informative on what domain features are invariant and shared among instances, so that the network sharpens its understanding and improves its semantic discriminative ability. Correspondingly, we also propose a novel "jury" mechanism, which is particularly effective in learning useful semantic feature commonalities among domains. Our complete model called STEAM can be interpreted as a novel probabilistic graphical model, for which the implementation requires convenient constructions of two kinds of memory banks: semantic feature bank and style feature bank. Empirical results show that our proposed framework surpasses the state-of-the-art methods by clear margins.

Lensless cameras provide a framework to build thin imaging systems by replacing the lens in a conventional camera with an amplitude or phase mask near the sensor. Existing methods for lensless imaging can recover the depth and intensity of the scene, but they require solving computationally-expensive inverse problems. Furthermore, existing methods struggle to recover dense scenes with large depth variations. In this paper, we propose a lensless imaging system that captures a small number of measurements using different patterns on a programmable mask. In this context, we make three contributions. First, we present a fast recovery algorithm to recover textures on a fixed number of depth planes in the scene. Second, we consider the mask design problem, for programmable lensless cameras, and provide a design template for optimizing the mask patterns with the goal of improving depth estimation. Third, we use a refinement network as a post-processing step to identify and remove artifacts in the reconstruction. These modifications are evaluated extensively with experimental results on a lensless camera prototype to showcase the performance benefits of the optimized masks and recovery algorithms over the state of the art.

The topical domain generalization (DG) problem asks trained models to perform well on an unseen target domain with different data statistics from the source training domains. In computer vision, data augmentation has proven one of the most effective ways of better exploiting the source data to improve domain generalization. However, existing approaches primarily rely on image-space data augmentation, which requires careful augmentation design, and provides limited diversity of augmented data. We argue that feature augmentation is a more promising direction for DG. We find that an extremely simple technique of perturbing the feature embedding with Gaussian noise during training leads to a classifier with domain-generalization performance comparable to existing state of the art. To model more meaningful statistics reflective of cross-domain variability, we further estimate the full class-conditional feature covariance matrix iteratively during training. Subsequent joint stochastic feature augmentation provides an effective domain randomization method, perturbing features in the directions of intra-class/cross-domain variability. We verify our proposed method on three standard domain generalization benchmarks, Digit-DG, VLCS and PACS, and show it is outperforming or comparable to the state of the art in all setups, together with experimental analysis to illustrate how our method works towards training a robust generalisable model.

We investigate the weakly-supervised scene graph generation, which is a challenging task since no correspondence of label and object is provided. The previous work regards such correspondence as a latent variable which is iteratively updated via nested optimization of the scene graph generation objective. However, we further reduce the complexity by decoupling it into an efficient first-order graph matching module optimized via contrastive learning to obtain such correspondence, which is used to train a standard scene graph generation model. The extensive experiments show that such a simple pipeline can significantly surpass the previous state-of-the-art by more than 30% on the Visual Genome dataset, both in terms of graph matching accuracy and scene graph quality. We believe this work serves as a strong baseline for future research.