ICCV 2023 论文列表

Image retrieval systems conventionally use a two-stage paradigm, leveraging global features for initial retrieval and local features for reranking. However, the scalability of this method is often limited due to the significant storage and computation cost incurred by local feature matching in the reranking stage. In this paper, we present SuperGlobal, a novel approach that exclusively employs global features for both stages, improving efficiency without sacrificing accuracy. SuperGlobal introduces key enhancements to the retrieval system, specifically focusing on the global feature extraction and reranking processes. For extraction, we identify sub-optimal performance when the widely-used ArcFace loss and Generalized Mean (GeM) pooling methods are combined and propose several new modules to improve GeM pooling. In the reranking stage, we introduce a novel method to update the global features of the query and top-ranked images by only considering feature refinement with a small set of images, thus being very compute and memory efficient. Our experiments demonstrate substantial improvements compared to the state of the art in standard benchmarks. Notably, on the Revisited Oxford+1M Hard dataset, our single-stage results improve by 7.1%, while our two-stage gain reaches 3.7% with a strong 64,865x speedup. Our two-stage system surpasses the current single-stage state-of-the-art by 16.3%, offering a scalable, accurate alternative for high-performing image retrieval systems with minimal time overhead. Code: https://github.com/ShihaoShao-GH/SuperGlobal.

Federated learning (FL) is a prevalent distributed machine learning approach that enables collaborative training of a global model across multiple devices without sharing local data. However, the presence of long-tailed data can negatively deteriorate the model's performance in real-world FL applications. Moreover, existing re-balance strategies are less effective for the federated long-tailed issue when directly utilizing local label distribution as the class prior at the clients' side. To this end, we propose a novel Global Balanced Multi-Expert (GBME) framework to optimize a balanced global objective, which does not require additional information beyond the standard FL pipeline. In particular, a proxy is derived from the accumulated gradients uploaded by the clients after local training, and is shared by all clients as the class prior for re-balance training. Such a proxy can also guide the client grouping to train a multi-expert model, where the knowledge from different clients can be aggregated via the ensemble of different experts corresponding to different client groups. To further strengthen the privacy-preserving ability, we present a GBME-p algorithm with a theoretical guarantee to prevent privacy leakage from the proxy. Extensive experiments on long-tailed decentralized datasets demonstrate the effectiveness of GBME and GBME-p, both of which show superior performance to state-of-the-art methods.

When applying a pre-trained 2D-to-3D Human Pose lifting model to a target unseen dataset, a large performance degradation is commonly encountered due to domain shift issues. We observe that the degradation is caused by two factors: 1) the large distribution gap over global positions of poses between the source and target datasets due to variant camera parameters and settings, and 2) the deficient diversity of local structures of poses in the training. To this end, we combine global adaptation and local generalization in PoseDA, a simple yet effective framework of unsupervised domain adaptation for 3D human pose estimation. Specifically, global adaptation aims to align global positions of poses from source domain to target domain with a proposed global position alignment (GPA) module. This module brings significant performance improvement without introducing additional learnable parameters. In addition, we propose local pose augmentation (LPA) to enhance the diversity of 3D poses following an adversarial training scheme consisting of 1) a augmentation generator that generates the parameters of pre-defined pose transformations and 2) an anchor discriminator to ensure the reality and quality of the augmented data. Our approach can be applicable to almost all 2D-3D lifting models. PoseDA achieves 61.3 mm of MPJPE on MPI-INF-3DHP under a cross-dataset evaluation setup, improving upon the previous state-of-the-art method by 10.2%.

Fast generation of high-quality 3D digital humans is important to a vast number of applications ranging from entertainment to professional concerns. Recent advances in differentiable rendering have enabled the training of 3D generative models without requiring 3D ground truths. However, the quality of the generated 3D humans still has much room to improve in terms of both fidelity and diversity. In this paper, we present Get3DHuman, a novel 3D human framework that can significantly boost the realism and diversity of the generated outcomes by only using a limited budget of 3D ground-truth data. Our key observation is that the 3D generator can profit from human-related priors learned through 2D human generators and 3D reconstructors. Specifically, we bridge the latent space of Get3DHuman with that of StyleGAN-Human via a specially-designed prior network, where the input latent code is mapped to the shape and texture feature volumes spanned by the pixel-aligned 3D reconstructor. The outcomes of the prior network are then leveraged as the supervisory signals for the main generator network. To ensure effective training, we further propose three tailored losses applied to the generated feature volumes and the intermediate feature maps. Extensive experiments demonstrate that Get3DHuman greatly outperforms the other state-of-the-art approaches and can support a wide range of applications including shape interpolation, shape re-texturing, and single-view reconstruction through latent inversion.

We generalize the class vectors found in neural networks to linear subspaces (i.e., points in the Grassmann manifold) and show that the Grassmann Class Representation (GCR) enables simultaneous improvement in accuracy and feature transferability. In GCR, each class is a subspace, and the logit is defined as the norm of the projection of a feature onto the class subspace. We integrate Riemannian SGD into deep learning frameworks such that class subspaces in a Grassmannian are jointly optimized with the rest model parameters. Compared to the vector form, the representative capability of subspaces is more powerful. We show that on ImageNet-1K, the top-1 errors of ResNet50-D, ResNeXt50, Swin-T, and Deit3-S are reduced by 5.6%, 4.5%, 3.0%, and 3.5%, respectively. Subspaces also provide freedom for features to vary, and we observed that the intra-class feature variability grows when the subspace dimension increases. Consequently, we found the quality of GCR features is better for downstream tasks. For ResNet50-D, the average linear transfer accuracy across 6 datasets improves from 77.98% to 79.70% compared to the strong baseline of vanilla softmax. For Swin-T, it improves from 81.5% to 83.4% and for Deit3, it improves from 73.8% to 81.4%. With these encouraging results, we believe that more applications could benefit from the Grassmann class representation. Code is released at https://github.com/innerlee/GCR.

In addition to color and textual information, geometry provides important cues for 3D scene reconstruction. However, current reconstruction methods only include geometry at the feature level thus not fully exploiting the geometric information. In contrast, this paper proposes a novel geometry integration mechanism for 3D scene reconstruction. Our approach incorporates 3D geometry at three levels, i.e. feature learning, feature fusion, and network supervision. First, geometry-guided feature learning encodes geometric priors to contain view-dependent information. Second, a geometry-guided adaptive feature fusion is introduced which utilizes the geometric priors as a guidance to adaptively generate weights for multiple views. Third, at the supervision level, taking the consistency between 2D and 3D normals into account, a consistent 3D normal loss is designed to add local constraints. Large-scale experiments are conducted on the ScanNet dataset, showing that volumetric methods with our geometry integration mechanism outperform state-of-the-art methods quantitatively as well as qualitatively. Volumetric methods with ours also show good generalization on the 7-Scenes and TUM RGB-D datasets.

For homography estimation, we propose Geometrized Transformer (GeoFormer), a new detector-free feature matching method. Current detector-free methods, e.g. LoFTR, lack an effective mean to accurately localize small and thus computationally feasible regions for cross-attention diffusion. We resolve the challenge with an extremely simple idea: using the classical RANSAC geometry for attentive region search. Given coarse matches by LoFTR, a homography is obtained with ease. Such a homography allows us to compute cross-attention in a focused manner, where key/value sets required by Transformers can be reduced to small fix-sized regions rather than an entire image. Local features can thus be enhanced by standard Transformers. We integrate GeoFormer into the LoFTR framework. By minimizing a multi-scale cross-entropy based matching loss on auto-generated training data, the network is trained in a fully self-supervised manner. Extensive experiments are conducted on multiple real-world datasets covering natural images, heavily manipulated pictures and retinal images. The proposed method compares favorably against the state-of-the-art.

Viewpoint is a fundamental modality that carries the interaction between observers and their environment. This paper proposes the first deep-learning framework for the viewpoint modality. The challenge in formulating learning frameworks for viewpoints resides in a suitable multimodal representation that links across the camera viewing space and 3D environment. Traditional approaches reduce the problem to image analysis instances, making them computationally expensive and not adequately modelling the intrinsic geometry and environmental context of 6DoF viewpoints. We improve these issues in two ways. 1) We propose a generalized viewpoint representation forgoing the analysis of photometric pixels in favor of encoded viewing ray embeddings attained from point cloud learning frameworks. 2) We propose a novel SE(3)-bijective 6D viewing ray, hyper-ray, that addresses the DoF deficiency problem of using 5DoF viewing rays representing 6DoF viewpoints. We demonstrate our approach has both efficiency and accuracy superiority over existing methods in novel real-world environments.

We present a learning-based method, namely GeoUDF, to tackle the long-standing and challenging problem of reconstructing a discrete surface from a sparse point cloud. To be specific, we propose a geometry-guided learning method for UDF and its gradient estimation that explicitly formulates the unsigned distance of a query point as the learnable affine averaging of its distances to the tangent planes of neighboring points on the surface. Besides, we model the local geometric structure of the input point clouds by explicitly learning a quadratic polynomial for each point. This not only facilitates upsampling the input sparse point cloud but also naturally induces unoriented normal, which further augments UDF estimation. Finally, to extract triangle meshes from the predicted UDF, we propose a customized edge-based marching cube module. We conduct extensive experiments and ablation studies to demonstrate the significant advantages of our method over state-of-the-art methods in terms of reconstruction accuracy, efficiency, and generality. The source code is publicly available at https://github.com/rsy6318/GeoUDF.

Multi-view camera-based 3D detection is a challenging problem in computer vision. Recent works leverage a pretrained LiDAR detection model to transfer knowledge to a camera-based student network. However, we argue that there is a major domain gap between the LiDAR BEV features and the camera-based BEV features, as they have different characteristics and are derived from different sources. In this paper, we propose Geometry Enhanced Masked Image Modeling (GeoMIM) to transfer the knowledge of the LiDAR model in a pretrain-finetune paradigm for improving the multi-view camera-based 3D detection. GeoMIM is a multi-camera vision transformer with Cross-View Attention (CVA) blocks that uses LiDAR BEV features encoded by the pretrained BEV model as learning targets. During pretraining, GeoMIM's decoder has a semantic branch completing dense perspective-view features and the other geometry branch reconstructing dense perspective-view depth maps. The depth branch is designed to be camera-aware by inputting the camera's parameters for better transfer capability. Extensive results demonstrate that GeoMIM outperforms existing methods on nuScenes benchmark, achieving state-of-the-art performance for camera-based 3D object detection and 3D segmentation.

Weakly supervised object localization (WSOL) remains challenging when learning object localization models from image category labels. Conventional methods that discriminatively train activation models ignore representative yet less discriminative object parts. In this study, we propose a generative prompt model (GenPromp), defining the first generative pipeline to localize less discriminative object parts by formulating WSOL as a conditional image denoising procedure. During training, GenPromp converts image category labels to learnable prompt embeddings which are fed to a generative model to conditionally recover the input image with noise and learn representative embeddings. During inference, GenPromp combines the representative embeddings with discriminative embeddings (queried from an off-the-shelf vision-language model) for both representative and discriminative capacity. The combined embeddings are finally used to generate multi-scale high-quality attention maps, which facilitate localizing full object extent. Experiments on CUB-200-2011 and ILSVRC show that GenPromp respectively outperforms the best discriminative models by 5.2% and 5.6% (Top-1 Loc), setting a solid baseline for WSOL with the generative model. Code is available at https://github.com/callsys/GenPromp.

We present a method to efficiently generate 3D-aware high-resolution images that are view-consistent across multiple target views. The proposed multiplane neural radiance model, named GMNR, consists of a novel a-guided view-dependent representation (a-VdR) module for learning view-dependent information. The a-VdR module, faciliated by an a-guided pixel sampling technique, computes the view-dependent representation efficiently by learning viewing direction and position coefficients. Moreover, we propose a view-consistency loss to enforce photometric similarity across multiple views. The GMNR model can generate 3D-aware high-resolution images that are view-consistent across multiple camera poses, while maintaining the computational efficiency in terms of both training and inference time. Experiments on three datasets demonstrate the effectiveness of the proposed modules, leading to favorable results in terms of both generation quality and inference time, compared to existing approaches. Our GMNR model generates 3D-aware images of 1024 x 1024 pixels with 17.6 FPS on a single V100. Code : https://github.com/VIROBO-15/GMNR

Federated learning has gained recognitions as a secure approach for safeguarding local private data in collaborative learning. But the advent of gradient inversion research has posed significant challenges to this premise by enabling a third-party to recover groundtruth images via gradients. While prior research has predominantly focused on low-resolution images and small batch sizes, this study highlights the feasibility of reconstructing complex images with high resolutions and large batch sizes. The success of the proposed method is contingent on constructing an over-parameterized convolutional network, so that images are generated before fitting to the gradient matching requirement. Practical experiments demonstrate that the proposed algorithm achieves high-fidelity image recovery, surpassing state-of-the-art competitors that commonly fail in more intricate scenarios. Consequently, our study shows that local participants in a federated learning system are vulnerable to potential data leakage issues. Source code is available at https://github.com/czhang024/CI-Net.

Skeleton-based action recognition has recently received considerable attention. Current approaches to skeleton-based action recognition are typically formulated as one-hot classification tasks and do not fully exploit the semantic relations between actions. For example, "make victory sign" and "thumb up" are two actions of hand gestures, whose major difference lies in the movement of hands. This information is agnostic from the categorical one-hot encoding of action classes but could be unveiled from the action description. Therefore, utilizing action description in training could potentially benefit representation learning. In this work, we propose a Generative Action-description Prompts (GAP) approach for skeleton-based action recognition. More specifically, we employ a pre-trained large-scale language model as the knowledge engine to automatically generate text descriptions for body parts movements of actions, and propose a multi-modal training scheme by utilizing the text encoder to generate feature vectors for different body parts and supervise the skeleton encoder for action representation learning. Experiments show that our proposed GAP method achieves noticeable improvements over various baseline models without extra computation cost at inference. GAP achieves new state-of-the-arts on popular skeleton-based action recognition benchmarks, including NTU RGB+D, NTU RGB+D 120 and NW-UCLA. The source code is available at https://github.com/MartinXM/GAP.

An emerging line of work has sought to generate plausible imagery from touch. Existing approaches, however, tackle only narrow aspects of the visuo-tactile synthesis problem, and lag significantly behind the quality of cross-modal synthesis methods in other domains. We draw on recent advances in latent diffusion to create a model for synthesizing images from tactile signals (and vice versa) and apply it to a number of visuo-tactile synthesis tasks. Using this model, we significantly outperform prior work on the tactile-driven stylization problem, i.e., manipulating an image to match a touch signal, and we are the first to successfully generate images from touch without additional sources of information about the scene. We also successfully use our model to address two novel synthesis problems: generating images that do not contain the touch sensor or the hand holding it, and estimating an image's shading from its reflectance and touch.

Representing wild sounds as images is an important but challenging task due to the lack of paired datasets between sound and image data and the significant differences in the characteristics of these two modalities. Previous studies have focused on generating images from sound in limited categories or music. In this paper, we propose a novel approach to generate images from wild sounds. First, we convert sound into text using audio captioning. Second, we propose audio attention and sentence attention to represent the rich characteristics of sound and visualize the sound. Lastly, we propose a direct sound optimization with CLIPscore and AudioCLIP and generate images with a diffusion-based model. In experiments, it shows that our model is able to generate high quality images from wild sounds and outperforms baselines in both quantitative and qualitative evaluations on wild audio datasets.

State-of-the-art lane detection methods often rely on specific knowledge about lanes -- such as straight lines and parametric curves -- to detect lane lines. While the specific knowledge can ease the modeling process, it poses challenges in handling lane lines with complex topologies (e.g., dense, forked, curved, etc.). Recently, dynamic convolution-based methods have shown promising performance by utilizing the features from some key locations of a lane line, such as the starting point, as convolutional kernels, and convoluting them with the whole feature map to detect lane lines. While such methods reduce the reliance on specific knowledge, the kernels computed from the key locations fail to capture the lane line's global structure due to its long and thin structure, leading to inaccurate detection of lane lines with complex topologies. In addition, the kernels resulting from the key locations are sensitive to occlusion and lane intersections. To overcome these limitations, we propose a transformer-based dynamic kernel generation architecture for lane detection. It utilizes a transformer to generate dynamic convolutional kernels for each lane line in the input image, and then detect these lane lines with dynamic convolution. Compared to the kernels generated from the key locations of a lane line, the kernels generated with the transformer can capture the lane line's global structure from the whole feature map, enabling them to effectively handle occlusions and lane lines with complex topologies. We evaluate our method on three lane detection benchmarks, and the results demonstrate its state-of-the-art performance. Specifically, our method achieves an F1 score of 63.40 on OpenLane and 88.47 on CurveLanes, surpassing the state of the art by 4.30 and 2.37 points, respectively.

Event-based motion deblurring has shown promising results by exploiting low-latency events. However, current approaches are limited in their practical usage, as they assume the same spatial resolution of inputs and specific blurriness distributions. This work addresses these limitations and aims to generalize the performance of event-based deblurring in real-world scenarios. We propose a scale-aware network that allows flexible input spatial scales and enables learning from different temporal scales of motion blur. A two-stage self-supervised learning scheme is then developed to fit real-world data distribution. By utilizing the relativity of blurriness, our approach efficiently ensures the restored brightness and structure of latent images and further generalizes deblurring performance to handle varying spatial and temporal scales of motion blur in a self-distillation manner. Our method is extensively evaluated, demonstrating remarkable performance, and we also introduce a real-world dataset consisting of multi-scale blurry frames and events to facilitate research in event-based deblurring.

A common architectural choice for deep metric learning is a convolutional neural network followed by global average pooling (GAP). Albeit simple, GAP is a highly effective way to aggregate information. One possible explanation for the effectiveness of GAP is considering each feature vector as representing a different semantic entity and GAP as a convex combination of them. Following this perspective, we generalize GAP and propose a learnable generalized sum pooling method (GSP). GSP improves GAP with two distinct abilities: i) the ability to choose a subset of semantic entities, effectively learning to ignore nuisance information, and ii) learning the weights corresponding to the importance of each entity. Formally, we propose an entropy-smoothed optimal transport problem and show that it is a strict generalization of GAP, i.e., a specific realization of the problem gives back GAP. We show that this optimization problem enjoys analytical gradients enabling us to use it as a direct learnable replacement for GAP. We further propose a zero-shot loss to ease the learning of GSP. We show the effectiveness of our method with extensive evaluations on 4 popular metric learning benchmarks. Code is available at: GSP-DML Framework