ICCV 2023 论文列表

Recently, DALL-E, a multimodal transformer language model, and its variants including diffusion models have shown high-quality text-to-image generation capabilities. However, despite the realistic image generation results, there has not been a detailed analysis of how to evaluate such models. In this work, we investigate the visual reasoning capabilities and social biases of different text-to-image models, covering both multimodal transformer language models and diffusion models. First, we measure three visual reasoning skills: object recognition, object counting, and spatial relation understanding. For this, we propose PaintSkills, a compositional diagnostic evaluation dataset that measures these skills. Despite the high-fidelity image generation capability, a large gap exists between the performance of recent models and the upper bound accuracy in object counting and spatial relation understanding skills. Second, we assess the gender and skin tone biases by measuring the gender/skin tone distribution of generated images across various professions and attributes. We demonstrate that recent text-to-image generation models learn specific biases about gender and skin tone from web image-text pairs. We hope our work will help guide future progress in improving text-to-image generation models on visual reasoning skills and learning socially unbiased representations.

Temporal sentence grounding (TSG) aims to locate a specific moment from an untrimmed video with a given natural language query. Recently, weakly supervised methods still have a large performance gap compared to fully supervised ones, while the latter requires laborious timestamp annotations. In this study, we aim to reduce the annotation cost yet keep competitive performance for TSG task compared to fully supervised ones. To achieve this goal, we investigate a recently proposed glance-supervised temporal sentence grounding task, which requires only single frame annotation (referred to as glance annotation) for each query. Under this setup, we propose a Dynamic Gaussian prior based Grounding framework with Glance annotation (D3G), which consists of a Semantic Alignment Group Contrastive Learning module (SA-GCL) and a Dynamic Gaussian prior Adjustment module (DGA). Specifically, SA-GCL samples reliable positive moments from a 2D temporal map via jointly leveraging Gaussian prior and semantic consistency, which contributes to aligning the positive sentence-moment pairs in the joint embedding space. Moreover, to alleviate the annotation bias resulting from glance annotation and model complex queries consisting of multiple events, we propose the DGA module, which adjusts the distribution dynamically to approximate the ground truth of target moments. Extensive experiments on three challenging benchmarks verify the effectiveness of the proposed D3G. It outperforms the state-of-the-art weakly supervised methods by a large margin and narrows the performance gap compared to fully supervised methods. Code is available at https://github.com/solicucu/D3G.

Realistic virtual humans play a crucial role in numerous industries, such as metaverse, intelligent healthcare, and self-driving simulation. But creating them on a large scale with high levels of realism remains a challenge. The utilization of deep implicit function sparks a new era of image-based 3D clothed human reconstruction, enabling pixel-aligned shape recovery with fine details. Subsequently, the vast majority of works locate the surface by regressing the deterministic implicit value for each point. However, should all points be treated equally regardless of their proximity to the surface? In this paper, we propose replacing the implicit value with an adaptive uncertainty distribution, to differentiate between points based on their distance to the surface. This simple "value to distribution" transition yields significant improvements on nearly all the baselines. Furthermore, qualitative results demonstrate that the models trained using our uncertainty distribution loss, can capture more intricate wrinkles, and realistic limbs. Code and models are available for research purposes at https://github.com/psyai-net/D-IF_release.

Recent progress in weakly supervised object detection is featured by a combination of multiple instance detection networks (MIDN) and ordinal online refinement. However, with only image-level annotation, MIDN inevitably assigns high scores to some unexpected region proposals when generating pseudo labels. These inaccurate high-scoring region proposals will mislead the training of subsequent refinement modules and thus hamper the detection performance. In this work, we explore how to ameliorate the quality of pseudo-labeling in MIDN. Formally, we devise Cyclic-Bootstrap Labeling (CBL), a novel weakly supervised object detection pipeline, which optimizes MIDN with rank information from a reliable teacher network. Specifically, we obtain this teacher network by introducing a weighted exponential moving average strategy to take advantage of various refinement modules. A novel class-specific ranking distillation algorithm is proposed to leverage the output of weighted ensembled teacher network for distilling MIDN with rank information. As a result, MIDN is guided to assign higher scores to accurate proposals, which further benefits final detection. Extensive experiments on the prevalent PASCAL VOC 2007 & 2012 and COCO datasets demonstrate the superior performance of our CBL framework.

Despite recent advances in 3D human mesh reconstruction, domain gap between training and test data is still a major challenge. Several prior works tackle the domain gap problem via test-time adaptation that fine-tunes a network relying on 2D evidence (e.g., 2D human keypoints) from test images. However, the high reliance on 2D evidence during adaptation causes two major issues. First, 2D evidence induces depth ambiguity, preventing the learning of accurate 3D human geometry. Second, 2D evidence is noisy or partially non-existent during test time, and such imperfect 2D evidence leads to erroneous adaptation. To overcome the above issues, we introduce CycleAdapt, which cyclically adapts two networks: a human mesh reconstruction network (HMRNet) and a human motion denoising network (MDNet), given a test video. In our framework, to alleviate high reliance on 2D evidence, we fully supervise HMRNet with generated 3D supervision targets by MDNet. Our cyclic adaptation scheme progressively elaborates the 3D supervision targets, which compensate for imperfect 2D evidence. As a result, our CycleAdapt achieves state-of-the-art performance compared to previous test-time adaptation methods. The codes are available in here: https://github.com/hygenie1228/CycleAdapt_RELEASE.

Coordinate networks are widely used in computer vision due to their ability to represent signals as compressed, continuous entities. However, training these networks with first-order optimizers can be slow, hindering their use in real-time applications. Recent works have opted for shallow voxel-based representations to achieve faster training, but this sacrifices memory efficiency. This work proposes a solution that leverages second-order optimization methods to significantly reduce training times for coordinate networks while maintaining their compressibility. Experiments demonstrate the effectiveness of this approach on various signal modalities, such as audio, images, videos, shape and neural radiance fields (NeRF).

Distilling knowledge from convolutional neural networks (CNNs) is a double-edged sword for vision transformers (ViTs). It boosts the performance since the image-friendly local-inductive bias of CNN helps ViT learn faster and better, but leading to two problems: (1) Network designs of CNN and ViT are completely different, which leads to different semantic levels of intermediate features, making spatial-wise knowledge transfer methods (e.g., feature mimicking) inefficient. (2) Distilling knowledge from CNN limits the network convergence in the later training period since ViT's capability of integrating global information is suppressed by CNN's local-inductive-bias supervision. To this end, we present Cumulative Spatial Knowledge Distillation (CSKD). CSKD distills spatial-wise knowledge to all patch tokens of ViT from the corresponding spatial responses of CNN, without introducing intermediate features. Furthermore, CSKD exploits a Cumulative Knowledge Fusion (CKF) module, which introduces the global response of CNN and increasingly emphasizes its importance during the training. Applying CKF leverages CNN's local inductive bias in the early training period and gives full play to ViT's global capability in the later one. Extensive experiments and analysis on ImageNet-1k and downstream datasets demonstrate the superiority of our CSKD. Code will be publicly available.

Medical image arbitrary-scale super-resolution (MIASSR) has recently gained widespread attention, aiming to supersample medical volumes at arbitrary scales via a single model. However, existing MIASSR methods face two major limitations: (i) reliance on high-resolution (HR) volumes and (ii) limited generalization ability, which restricts their applications in various scenarios. To overcome these limitations, we propose Cube-based Neural Radiance Field (CuNeRF), a zero-shot MIASSR framework that is able to yield medical images at arbitrary scales and free viewpoints in a continuous domain. Unlike existing MISR methods that only fit the mapping between low-resolution (LR) and HR volumes, CuNeRF focuses on building a continuous volumetric representation from each LR volume without the knowledge from the corresponding HR one. This is achieved by the proposed differentiable modules: cube-based sampling, isotropic volume rendering, and cube-based hierarchical rendering. Through extensive experiments on magnetic resource imaging (MRI) and computed tomography (CT) modalities, we demonstrate that CuNeRF can synthesize high-quality SR medical images, which outperforms state-of-the-art MISR methods, achieving better visual verisimilitude and fewer objectionable artifacts. Compared to existing MISR methods, our CuNeRF is more applicable in practice.

Source-free domain adaptive semantic segmentation has gained increasing attention recently. It eases the requirement of full data access to the source domain by transferring knowledge only from a well-trained source model. However, reducing the uncertainty of the target pseudo labels becomes inevitably more challenging without the supervision of the labeled source data. In this work, we propose a novel asymmetric two-stream architecture that learns more robustly from noisy pseudo labels. Our approach simultaneously conducts dual-head pseudo label denoising and cross-modal consistency regularization. Towards the former, we introduce a multimodal auxiliary network during training (and discard it during inference), which effectively enhances the pseudo labels' correctness by leveraging the guidance from the depth information. Towards the latter, we enforce a new cross-modal pixel-wise consistency between the predictions of the two streams, encouraging our model to behave smoothly for both modality variance and image perturbations. It serves as an effective regularization to further reduce the impact of the inaccurate pseudo labels in source-free unsupervised domain adaptation. Experiments on GTA5 to Cityscapes and SYNTHIA to Cityscapes benchmarks demonstrate the superiority of our proposed method, obtaining the new state-of-the-art mIoU of 57.7% and 57.5%, respectively.

We present CrossLoc3D, a novel 3D place recognition method that solves a large-scale point matching problem in a cross-source setting. Cross-source point cloud data corresponds to point sets captured by depth sensors with different accuracies or from different distances and perspectives. We address the challenges in terms of developing 3D place recognition methods that account for the representation gap between points captured by different sources. Our method handles cross-source data by utilizing multi-grained features and selecting convolution kernel sizes that correspond to most prominent features. Inspired by the diffusion models, our method uses a novel iterative refinement process that gradually shifts the embedding spaces from different sources to a single canonical space for better metric learning. In addition, we present CS-Campus3D, the first 3D aerial-ground cross-source dataset consisting of point cloud data from both aerial and ground LiDAR scans. The point clouds in CS-Campus3D have representation gaps and other features like different views, point densities, and noise patterns. We show that our CrossLoc3D algorithm can achieve an improvement of 4.74% - 15.37% in terms of the top 1 average recall on our CS-Campus3D benchmark and achieves performance comparable to state-of-the-art 3D place recognition method on the Oxford RobotCar. We will release the code and CS-Campus3D benchmark.

Multi-view clustering aims to extract valuable information from different sources or perspectives. Over the years, the deep neural network has demonstrated its superior representation learning capability in multi-view clustering and achieved impressive performance. However, most existing deep clustering approaches are dedicated to merging and exploring the consistent latent representation across multiple views while overlooking the abundant complementary information in each view. Furthermore, finding correlations between multiple views in an unsupervised setting is a significant challenge. To tackle these issues, we present a novel Cross-view Topology based Consistent and Complementary information extraction framework, termed CTCC. In detail, deep embedding can be obtained from the bipartite graph learning module for each view individually. CTCC then constructs the cross-view topological graph based on the OT distance between the bipartite graph of each view. Utilizing the above graph, we maximize the mutual information across views to learn consistent information and enhance the complementarity of each view by selectively isolating distributions from each other. Extensive experiments on five challenging datasets verify that CTCC outperforms existing methods significantly.

Live commerce is the act of selling products online through livestreaming. The customer's diverse demands for online products introduces more challenges to Livestreaming Product Recognition. Previous works are either focus on fashion clothing data or subject to single-modal input, thus inconsistent with the real-world scenario where multimodal data from various categories are present. In this paper, we contribute LPR4M, a large-scale multimodal dataset that covers 34 categories, comprises 3 modalities (image, video, and text), and is 50 times larger than the largest publicly available dataset. In addition, LPR4M contains diverse videos and noise modality pair while also having a long-tailed distribution, resembling real-world problems. Moreover, a cRoss-vIew semantiC alignmEnt (RICE) model is proposed to learn discriminative instance features from the two views (image and video) of products via instance-level contrastive learning as well as cross-view patch-level feature propagation. A novel Patch Feature Reconstruction loss is proposed to penalize the semantic misalignment between the cross-view patches. Extensive ablation studies demonstrate the effectiveness of RICE and provide insights into the importance of dataset diversity and expressivity.

Hierarchical clustering is a natural approach to discover ontologies from data. Yet, existing approaches are hampered by their inability to scale to large datasets and the discrete encoding of the hierarchy. We introduce scalable Hyperbolic Hierarchical Clustering (sHHC) which overcomes these limitations by learning continuous hierarchies in hyperbolic space. Our hierarchical clustering is of high quality and can be obtained in a fraction of the runtime. Additionally, we demonstrate the strength of sHHC on a downstream cross-modal self-supervision task. By using the discovered hierarchies from sound and vision to construct continuous hierarchical pseudo-labels we can efficiently optimize a network for activity recognition and obtain competitive performance compared to recent self-supervised learning models. Our findings demonstrate the strength of Hyperbolic Hierarchical Clustering and its potential for Self-Supervised Learning.

We present a novel method for automatic vectorized avatar generation from a single portrait image. Most existing approaches that create avatars rely on image-to-image translation methods, which present some limitations when applied to 3D rendering, animation, or video. Instead, we leverage modality-specific autoencoders trained on large-scale unpaired portraits and parametric avatars, and then learn a mapping between both modalities via an alignment module trained on a significantly smaller amount of data. The resulting cross-modal latent space preserves facial identity, producing more visually appealing and higher fidelity avatars than previous methods, as supported by our quantitative and qualitative evaluations. Moreover, our method's virtue of being resolution-independent makes it highly versatile and applicable in a wide range of settings.

Neural Radiance Fields (NeRF) is a revolutionary approach for rendering scenes by sampling a single ray per pixel and it has demonstrated impressive capabilities in novel-view synthesis from static scene images. However, in practice, we usually need to recover NeRF from unconstrained image collections, which poses two challenges: 1) the images often have dynamic changes in appearance because of different capturing time and camera settings; 2) the images may contain transient objects such as humans and cars, leading to occlusion and ghosting artifacts. Conventional approaches seek to address these challenges by locally utilizing a single ray to synthesize a color of a pixel. In contrast, humans typically perceive appearance and objects by globally utilizing information across multiple pixels. To mimic the perception process of humans, in this paper, we propose Cross-Ray NeRF (CR-NeRF) that leverages interactive information across multiple rays to synthesize occlusion-free novel views with the same appearances as the images. Specifically, to model varying appearances, we first propose to represent multiple rays with a novel cross-ray feature and then recover the appearance by fusing global statistics, i.e., feature covariance of the rays and the image appearance. Moreover, to avoid occlusion introduced by transient objects, we propose a transient objects handler and introduce a grid sampling strategy for masking out the transient objects. We theoretically find that leveraging correlation across multiple rays promotes capturing more global information. Moreover, extensive experimental results on large real-world datasets verify the effectiveness of CR-NeRF.

With the rapid advances in high-throughput sequencing technologies, the focus of survival analysis has shifted from examining clinical indicators to incorporating genomic profiles with pathological images. However, existing methods either directly adopt a straightforward fusion of pathological features and genomic profiles for survival prediction, or take genomic profiles as guidance to integrate the features of pathological images. The former would overlook intrinsic cross-modal correlations. The latter would discard pathological information irrelevant to gene expression. To address these issues, we present a Cross-Modal Translation and Alignment (CMTA) framework to explore the intrinsic cross-modal correlations and transfer potential complementary information. Specifically, we construct two parallel encoder-decoder structures for multi-modal data to integrate intra-modal information and generate cross-modal representation. Taking the generated cross-modal representation to enhance and recalibrate intra-modal representation can significantly improve its discrimination for comprehensive survival analysis. To explore the intrinsic cross-modal correlations, we further design a cross-modal attention module as the information bridge between different modalities to perform cross-modal interactions and transfer complementary information. Our extensive experiments on five public TCGA datasets demonstrate that our proposed framework outperforms the state-of-the-art methods. The source code has been released.

This paper addresses the problem of cross-modal object tracking from RGB videos and event data. Rather than constructing a complex cross-modal fusion network, we explore the great potential of a pre-trained vision Transformer (ViT). Particularly, we delicately investigate plug-and-play training augmentations that encourage the ViT to bridge the vast distribution gap between the two modalities, enabling comprehensive cross-modal information interaction and thus enhancing its ability. Specifically, we propose a mask modeling strategy that randomly masks a specific modality of some tokens to enforce the interaction between tokens from different modalities interacting proactively. To mitigate network oscillations resulting from the masking strategy and further amplify its positive effect, we then theoretically propose an orthogonal high-rank loss to regularize the attention matrix. Extensive experiments demonstrate that our plug-and-play training augmentation techniques can significantly boost state-of-the-art one-stream and two-stream trackers to a large extent in terms of both tracking precision and success rate. Our new perspective and findings will potentially bring insights to the field of leveraging powerful pre-trained ViTs to model cross-modal data. The code is publicly available at https://github.com/ZHU-Zhiyu/High-Rank_RGB-Event_Tracker.

An important challenge in vision-based action recognition is the embedding of spatiotemporal features with two or more heterogeneous modalities into a single feature. In this study, we propose a new 3D deformable transformer for action recognition with adaptive spatiotemporal receptive fields and a cross-modal learning scheme. The 3D deformable transformer consists of three attention modules: 3D deformability, local joint stride, and temporal stride attention. The two cross-modal tokens are input into the 3D deformable attention module to create a cross-attention token with a reflected spatiotemporal correlation. Local joint stride attention is applied to spatially combine attention and pose tokens. Temporal stride attention temporally reduces the number of input tokens in the attention module and supports temporal expression learning without the simultaneous use of all tokens. The deformable transformer iterates L-times and combines the last cross-modal token for classification. The proposed 3D deformable transformer was tested on the NTU60, NTU120, FineGYM, and PennAction datasets, and showed results better than or similar to pre-trained state-of-the-art methods even without a pre-training process. In addition, by visualizing important joints and correlations during action recognition through spatial joint and temporal stride attention, the possibility of achieving an explainable potential for action recognition is presented.

The proliferation of short video and live-streaming platforms has revolutionized how consumers engage in online shopping. Instead of browsing product pages, consumers are now turning to rich-content e-commerce, where they can purchase products through dynamic and interactive media like short videos and live streams. This emerging form of online shopping has presented new opportunities for platforms to enhance user engagement and shopping experience. However, it has also introduced technical challenges, as products may be presented differently across various media domains. Therefore, a unified product representation is essential for achieving cross-domain product recognition to ensure an optimal user search experience and effective product recommendations. Despite the urgent industrial need for a unified cross-domain product representation, previous studies have predominantly focused only on product pages without taking into account short videos and live streams. To fill the gap in the rich-content e-commerce area, in this paper, we introduce a large-scale cross-domain poduct recognition dataset, called ROPE. ROPE covers a wide range of product categories and contains over 180,000 products, corresponding to millions of short videos and live streams. It is the first dataset to cover product pages, short videos, and live streams simultaneously, providing the basis for establishing a unified product representation across different media domains. Furthermore, we propose a cross-domain product representation framework, namely COPE, which unifies product representations in different domains through multimodal learning including text and vision. Extensive experiments on downstream tasks like cross-modal retrieval and classification demonstrate the effectiveness of COPE in learning a joint feature space for all product domains.