ICCV 2021 Papers

Video data is with complex temporal dynamics due to various factors such as camera motion, speed variation, and different activities. To effectively capture this diverse motion pattern, this paper presents a new temporal adaptive module (TAM) to generate video-specific temporal kernels based on its own feature map. TAM proposes a unique two-level adaptive modeling scheme by decoupling the dynamic kernel into a location sensitive importance map and a location invariant aggregation weight. The importance map is learned in a local temporal window to capture short-term information, while the aggregation weight is generated from a global view with a focus on long-term structure. TAM is a modular block and could be integrated into 2D CNNs to yield a powerful video architecture (TANet) with a very small extra computational cost. The extensive experiments on Kinetics-400 and Something-Something datasets demonstrate that our TAM outperforms other temporal modeling methods consistently, and achieves the state-of-the-art performance under the similar complexity. The code is available at https://github.com/liu-zhy/temporal-adaptive-module.

Contrastive learning has been widely used to train transformer-based vision-language models for video-text alignment and multi-modal representation learning. This paper presents a new algorithm called Token-Aware Cascade contrastive learning (TACo) that improves contrastive learning using two novel techniques. The first is the token-aware contrastive loss which is computed by taking into account the syntactic classes of words. This is motivated by the observation that for a video-text pair, the content words in the text, such as nouns and verbs, are more likely to be aligned with the visual contents in the video than the function words. Second, a cascade sampling method is applied to generate a small set of hard negative examples for efficient loss estimation for multi-modal fusion layers. To validate the effectiveness of TACo, in our experiments we finetune pretrained models for a set of downstream tasks including text-video retrieval (YouCook2, MSR-VTT and ActivityNet), video action step localization (CrossTask), video action segmentation (COIN). Our results show that our models attain consistent improvements across different experimental settings over previous methods, setting new state-of-the-art on three public text-video retrieval benchmarks of YouCook2, MSR-VTT and ActivityNet.

Most existing domain adaptation methods focus on adaptation from only one source domain, however, in practice there are a number of relevant sources that could be leveraged to help improve performance on target domain. We propose a novel approach named T-SVDNet to address the task of Multi-source Domain Adaptation (MDA), which is featured by incorporating Tensor Singular Value Decomposition (T-SVD) into a neural network's training pipeline. Overall, high-order correlations among multiple domains are fully explored so as to better bridge the domain gap in this work. Specifically, we impose Tensor-Low-Rank (TLR) constraint on the tensor obtained by stacking up a group of prototypical similarity matrices, aiming at capturing consistent data structure across different domains. Furthermore, to avoid negative transfer brought by noisy source data, we propose a novel uncertainty-aware weighting strategy to adaptively assign weights to different source domains and samples based on the result of uncertainty estimation. Extensive experiments conducted on public benchmarks demonstrate the superiority of our model in addressing the task of MDA compared to state-of-the-art methods.

We develop a conceptually simple, flexible, and effective framework (named T-Net) for two-view correspondence learning. Given a set of putative correspondences, we reject outliers and regress the relative pose encoded by the essential matrix, by an end-to-end framework, which is consisted of two novel structures: "-" structure and "|" structure. "-" structure adopts an iterative strategy to learn correspondence features. "|" structure integrates all the features of the iterations and outputs the correspondence weight. In addition, we introduce Permutation-Equivariant Context Squeeze-and-Excitation module, an adapted version of SE module, to process sparse correspondences in a permutation-equivariant way and capture both global and channel-wise contextual information. Extensive experiments on outdoor and indoor scenes show that the proposed T-Net achieves state-of-the-art performance. On outdoor scenes (YFCC100M dataset), T-Net achieves an mAP of 52.28%, a 34.22% precision increase from the best-published result (38.95%). On indoor scenes (SUN3D dataset), T-Net (19.71%) obtains a 21.82% precision increase from the best-published result (16.18%).

Lesion segmentation in medical imaging has been an important topic in clinical research. Researchers have proposed various detection and segmentation algorithms to address this task. Recently, deep learning-based approaches have significantly improved the performance over conventional methods. However, most state-of-the-art deep learning methods require the manual design of multiple network components and training strategies. In this paper, we propose a new automated machine learning algorithm, T-AutoML, which not only searches for the best neural architecture, but also finds the best combination of hyper-parameters and data augmentation strategies simultaneously. The proposed method utilizes the modern transformer model, which is introduced to adapt to the dynamic length of the search space embedding and can significantly improve the ability of the search. We validate T-AutoML on several large-scale public lesion segmentation data-sets and achieve state-of-the-art performance.

Few-shot class incremental learning (FSCIL) aims to incrementally add sets of novel classes to a well-trained base model in multiple training sessions with the restriction that only a few novel instances are available per class. While learning novel classes, FSCIL methods gradually forget base (old) class training and overfit to a few novel class samples. Existing approaches have addressed this problem by computing the class prototypes from the visual or semantic word vector domain. In this paper, we propose addressing this problem using a mixture of subspaces. Subspaces define the cluster structure of the visual domain and help to describe the visual and semantic domain considering the overall distribution of the data. Additionally, we propose to employ a variational autoencoder (VAE) to generate synthesized visual samples for augmenting pseudo-feature while learning novel classes incrementally. The combined effect of the mixture of subspaces and synthesized features reduces the forgetting and overfitting problem of FSCIL. Extensive experiments on three image classification datasets show that our proposed method achieves competitive results compared to state-of-the-art methods.

How can we animate 3D-characters from a movie script or move robots by simply telling them what we would like them to do?" How unstructured and complex can we make a sentence and still generate plausible movements from it?" These are questions that need to be answered in the long-run, as the field is still in its infancy. Inspired by these problems, we present a new technique for generating compositional actions, which handles complex input sentences. Our output is a 3D pose sequence depicting the actions in the input sentence. We propose a hierarchical two-stream sequential model to explore a finer joint-level mapping between natural language sentences and 3D pose sequences corresponding to the given motion. We learn two manifold representations of the motion, one each for the upper body and the lower body movements. Our model can generate plausible pose sequences for short sentences describing single actions as well as long complex sentences describing multiple sequential and compositional actions. We evaluate our proposed model on the publicly available KIT Motion-Language Dataset containing 3D pose data with human-annotated sentences. Experimental results show that our model advances the state-of-the-art on text-based motion synthesis in objective evaluations by a margin of 50%. Qualitative evaluations based on a user study indicate that our synthesized motions are perceived to be the closest to the ground-truth motion captures for both short and compositional sentences.

Visible infrared person re-identification (VI-REID) aims to match pedestrian images between the daytime visible and nighttime infrared camera views. The large cross-modality discrepancies have become the bottleneck which limits the performance of VI-REID. Existing methods mainly focus on capturing cross-modality sharable representations by learning an identity classifier. However, the heterogeneous pedestrian images taken by different spectrum cameras differ significantly in image styles, resulting in inferior discriminability of feature representations. To alleviate the above problem, this paper explores the correlation between two modalities and proposes a novel syncretic modality collaborative learning (SMCL) model to bridge the cross-modality gap. A new modality that incorporates features of heterogeneous images is constructed automatically to steer the generation of modality-invariant representations. Challenge enhanced homogeneity learning (CEHL) and auxiliary distributional similarity learning (ADSL) are integrated to project heterogeneous features on a unified space and enlarge the inter-class disparity, thus strengthening the discriminative power. Extensive experiments on two cross-modality benchmarks demonstrate the effectiveness and superiority of the proposed method. Especially, on SYSU-MM01 dataset, our SMCL model achieves 67.39% rank-1 accuracy and 61.78% mAP, surpassing the cutting-edge works by a large margin.

Synchronization refers to the problem of inferring the unknown values attached to vertices of a graph where edges are labelled with the ratio of the incident vertices, and labels belong to a group. This paper addresses the synchronization problem on multi-graphs, that are graphs with more than one edge connecting the same pair of nodes. The problem naturally arises when multiple measures are available to model the relationship between two vertices. This happens when different sensors measure the same quantity, or when the original graph is partitioned into sub-graphs that are solved independently. In this case, the relationships among sub-graphs give rise to multi-edges and the problem can be traced back to a multi-graph synchronization. The baseline solution reduces multi-graphs to simple ones by averaging their multi-edges, however this approach falls short because: i) averaging is well defined only for some groups and ii) the resulting estimator is less precise and accurate, as we prove empirically. Specifically, we present MultiSynch, a synchronization algorithm for multi-graphs that is based on a principled constrained eigenvalue optimization. MultiSynch is a general solution that can cope with any linear group and we show to be profitably usable both on synthetic and real problems.

With the recent success of deep neural networks, remarkable progress has been achieved on face recognition. However, collecting large-scale real-world training data for face recognition has turned out to be challenging, especially due to the label noise and privacy issues. Meanwhile, existing face recognition datasets are usually collected from web images, lacking detailed annotations on attributes (e.g., pose and expression), so the influences of different attributes on face recognition have been poorly investigated. In this paper, we address the above-mentioned issues in face recognition using synthetic face images, i.e., SynFace. Specifically, we first explore the performance gap between recent state-of-the-art face recognition models trained with synthetic and real face images. We then analyze the underlying causes behind the performance gap, e.g., the poor intra-class variations and the domain gap between synthetic and real face images. Inspired by this, we devise the SynFace with identity mixup (IM) and domain mixup (DM) to mitigate the above performance gap, demonstrating the great potentials of synthetic data for face recognition. Furthermore, with the controllable face synthesis model, we can easily manage different factors of synthetic face generation, including pose, expression, illumination, the number of identities, and samples per identity. Therefore, we also perform a systematically empirical analysis on synthetic face images to provide some insights on how to effectively utilize synthetic data for face recognition.

Optimizing the K-class hyperplanes in the latent space has become the standard paradigm for efficient representation learning. However, it's almost impossible to find an optimal K-class hyperplane to accurately describe the latent space of massive noisy data. For this potential problem, we constructively propose a new method, named Switchable K-class Hyperplanes (SKH), to sufficiently describe the latent space by the mixture of K-class hyperplanes. It can directly replace the conventional single K-class hyperplane optimization as the new paradigm for noise-robust representation learning. When collaborated with the popular ArcFace on million-level data representation learning, we found that the switchable manner in SKH can effectively eliminate the gradient conflict generated by real-world label noise on a single K-class hyperplane. Moreover, combined with the margin-based loss functions (e.g. ArcFace), we propose a simple Posterior Data Clean strategy to reduce the model optimization deviation on clean dataset caused by the reduction of valid categories in each K-class hyperplane. Extensive experiments demonstrate that the proposed SKH easily achieves new state-of-the-art on IJB-B and IJB-C by encouraging noise-robust representation learning.

This paper presents a new vision Transformer, called Swin Transformer, that capably serves as a general-purpose backbone for computer vision. Challenges in adapting Transformer from language to vision arise from differences between the two domains, such as large variations in the scale of visual entities and the high resolution of pixels in images compared to words in text. To address these differences, we propose a hierarchical Transformer whose representation is computed with Shifted windows. The shifted windowing scheme brings greater efficiency by limiting self-attention computation to non-overlapping local windows while also allowing for cross-window connection. This hierarchical architecture has the flexibility to model at various scales and has linear computational complexity with respect to image size. These qualities of Swin Transformer make it compatible with a broad range of vision tasks, including image classification (87.3 top-1 accuracy on ImageNet-1K) and dense prediction tasks such as object detection (58.7 box AP and 51.1 mask AP on COCO test-dev) and semantic segmentation (53.5 mIoU on ADE20K val). Its performance surpasses the previous state-of-the-art by a large margin of +2.7 box AP and +2.6 mask AP on COCO, and +3.2 mIoU on ADE20K, demonstrating the potential of Transformer-based models as vision backbones. The hierarchical design and the shifted window approach also prove beneficial for all-MLP architectures. The code and models are publicly available at https://github.com/microsoft/Swin-Transformer.

In this paper we present SurfaceNet, an approach for estimating spatially-varying bidirectional reflectance distribution function (SVBRDF) material properties from a single image. We pose the problem as an image translation task and propose a novel patch-based generative adversarial network (GAN) that is able to produce high-quality, high-resolution surface reflectance maps. The employment of the GAN paradigm has a twofold objective: 1) allowing the model to recover finer details than standard translation models; 2) reducing the domain shift between synthetic and real data distributions in an unsupervised way. An extensive evaluation, carried out on a public benchmark of synthetic and real images under different illumination conditions, shows that SurfaceNet largely outperforms existing SVBRDF reconstruction methods, both quantitatively and qualitatively. Furthermore, SurfaceNet exhibits a re-markable ability in generating high-quality maps from real samples without any supervision at training time.

Recent advances in deep generative models have led to immense progress in 3D shape synthesis. While existing models are able to synthesize shapes represented as voxels, point-clouds, or implicit functions, these methods only indirectly enforce the plausibility of the final 3D shape surface. Here we present a 3D shape synthesis framework (SurfGen) that directly applies adversarial training to the object surface. Our approach uses a differentiable spherical projection layer to capture and represent the explicit zero isosurface of an implicit 3D generator as functions defined on the unit sphere. By processing the spherical representation of 3D object surfaces with a spherical CNN in an adversarial setting, our generator can better learn the statistics of natural shape surfaces. We evaluate our model on large-scale shape datasets, and demonstrate that the end-to-end trained model is capable of generating high fidelity 3D shapes with diverse topology

Current approaches for video grounding propose kinds of complex architectures to capture the video-text relations, and have achieved impressive improvements. However, it is hard to learn the complicated multi-modal relations by only architecture designing in fact. In this paper, we introduce a novel Support-set Based Cross-Supervision (Sscs) module which can improve existing methods during training phase without extra inference cost. The contrastive objective aims to learn effective representations by contrastive learning, while the caption objective can train a powerful video encoder supervised by texts. Due to the co-existence of some visual entities in both ground-truth and background intervals, i.e., mutual exclusion, naively contrastive learning is unsuitable to video grounding. We address the problem by boosting the cross-supervision with the support-set concept, which collects visual information from the whole video and eliminates the mutual exclusion of entities. Combined with the original objective, Sscs can enhance the abilities of multi-modal relation modeling for existing approaches. We extensively evaluate Sscs on three challenging datasets, and show that our method can improves current state-of-the-art methods by large margins, especially 6.35% in terms of R1@0.5 on Charades-STA.

Superpoints are formed by grouping similar points with local geometric structures, which can effectively reduce the number of primitives of point clouds for subsequent point cloud processing. Existing superpoint methods mainly focus on employing clustering or graph partition to generate superpoints with handcrafted or learned features. Nonetheless, these methods cannot learn superpoints of point clouds with an end-to-end network. In this paper, we develop a new deep iterative clustering network to directly generate superpoints from irregular 3D point clouds in an end-to-end manner. Specifically, in our clustering network, we first jointly learn a soft point-superpoint association map from the coordinate and feature spaces of point clouds, where each point is assigned to the superpoint with a learned weight. Furthermore, we then iteratively update the association map and superpoint centers so that we can more accurately group the points into the corresponding superpoints with locally similar geometric structures. Finally, by predicting the pseudo labels of the superpoint centers, we formulate a label consistency loss on the points and superpoint centers to train the network. Extensive experiments on various datasets indicate that our method not only achieves the state-of-the-art on superpoint generation but also improves the performance of point cloud semantic segmentation. Code is available at https://github.com/fpthink/SPNet.

Conventional face super-resolution methods usually assume testing low-resolution (LR) images lie in the same domain as the training ones. Due to different lighting conditions and imaging hardware, domain gaps between training and testing images inevitably occur in many real-world scenarios. Neglecting those domain gaps would lead to inferior face super-resolution (FSR) performance. However, how to transfer a trained FSR model to a target domain efficiently and effectively has not been investigated. To tackle this problem, we develop a Domain-Aware Pyramid-based Face Super-Resolution network, named DAP-FSR network. Our DAP-FSR is the first attempt to super-resolve LR faces from a target domain by exploiting only a pair of high-resolution (HR) and LR exemplar in the target domain. To be specific, our DAP-FSR firstly employs its encoder to extract the multi-scale latent representations of the input LR face. Considering only one target domain example is available, we propose to augment the target domain data by mixing the latent representations of the target domain face and source domain ones and then feed the mixed representations to the decoder of our DAP-FSR. The decoder will generate new face images resembling the target domain image style. The generated HR faces in turn are used to optimize our decoder to reduce the domain gap. By iteratively updating the latent representations and our decoder, our DAP-FSR will be adapted to the target domain, thus achieving authentic and high-quality upsampled HR faces. Extensive experiments on three benchmarks validate the effectiveness and superior performance of our DAP-FSR compared to the state-of-the-art methods.

Recently, a novel retina-inspired camera, namely spike camera, has shown great potential for recording high-speed dynamic scenes. Unlike the conventional digital cameras that compact the visual information within the exposure interval into a single snapshot, the spike camera continuously outputs binary spike streams to record the dynamic scenes, yielding a very high temporal resolution. Most of the existing reconstruction methods for spike camera focus on reconstructing images with the same resolution as spike camera. However, as a trade-off of high temporal resolution, the spatial resolution of spike camera is limited, resulting in inferior details of the reconstruction. To address this issue, we develop a spike camera super-resolution framework, aiming to super resolve high-resolution intensity images from the low-resolution binary spike streams. Due to the relative motion between the camera and the objects to capture, the spikes fired by the same sensor pixel no longer describes the same points in the external scene. In this paper, we properly exploit the relative motion and derive the relationship between light intensity and each spike, so as to recover the external scene with both high temporal and high spatial resolution. Experimental results demonstrate that the proposed method can reconstruct pleasant high-resolution images from low-resolution spike streams.

Humans perform co-saliency detection by first summarizing the consensus knowledge in the whole group and then searching corresponding objects in each image. Previous methods usually lack robustness, scalability, or stability for the first process and simply fuse consensus features with image features for the second process. In this paper, we propose a novel consensus-aware dynamic convolution model to explicitly and effectively perform the "summarize and search" process. To summarize consensus image features, we first summarize robust features for every single image using an effective pooling method and then aggregate cross-image consensus cues via the self-attention mechanism. By doing this, our model meets the scalability and stability requirements. Next, we generate dynamic kernels from consensus features to encode the summarized consensus knowledge. Two kinds of kernels are generated in a supplementary way to summarize fine-grained image-specific consensus object cues and the coarse group-wise common knowledge, respectively. Then, we can effectively perform object searching by employing dynamic convolution at multiple scales. Besides, a novel and effective data synthesis method is also proposed to train our network. Experimental results on four benchmark datasets verify the effectiveness of our proposed method. Our code and saliency maps are available at https://github.com/nnizhang/CADC.