论文检索

Mengqi Xue, Qihan Huang, Haofei Zhang, Jingwen Hu, Jie Song, Mingli Song, Canghong Jin

Prototypical part network (ProtoPNet) and its variants have drawn wide attention and been applied to various tasks due to their inherent self-explanatory property. Previous ProtoPNets are primarily built upon convolutional neural networks (CNNs). Therefore, it is natural to investigate whether these explainable methods can be advantageous for the recently emerged Vision Transformers (ViTs). However, directly utilizing ViT-backed models as backbones can lead to prototypes paying excessive attention to background positions rather than foreground objects (i.e., the “distraction” problem). To address the problem, this paper proposes prototypical part Transformer (ProtoPFormer) for interpretable image recognition. Based the architectural characteristics of ViTs, we modify the original ProtoPNet by creating separate global and local branches, each accompanied by corresponding prototypes that can capture and highlight representative holistic and partial features. Specifically, the global prototypes can guide local prototypes to concentrate on the foreground and effectively suppress the background influence. Subsequently, local prototypes are explicitly supervised to concentrate on different discriminative visual parts. Finally, the two branches mutually correct each other and jointly make the final decisions. Moreover, extensive experiments demonstrate that ProtoPFormer can consistently achieve superior performance on accuracy, visualization results, and quantitative interpretability evaluation over the state-of-the-art (SOTA) baselines. Our code has been released at https://github.com/zju-vipa/ProtoPFormer.

Yuanwu Xu, Feifan Song, Haofeng Zhang

Few-shot object counting aims to count the number of objects in a query image that belong to the same class as the given exemplar images. Existing methods compute the similarity between the query image and exemplars in the 2D spatial domain and perform regression to obtain the counting number. However, these methods overlook the rich information about the spatial distribution of similarity on the exemplar images, leading to significant impact on matching accuracy. To address this issue, we propose a network learning Spatial Similarity Distribution (SSD) for few-shot object counting, which preserves the spatial structure of exemplar features and calculates a 4D similarity pyramid point-to-point between the query features and exemplar features, capturing the complete distribution information for each point in the 4D similarity space. We propose a Similarity Learning Module (SLM) which applies the efficient center-pivot 4D convolutions on the similarity pyramid to map different similarity distributions to distinct predicted density values, thereby obtaining accurate count. Furthermore, we also introduce a Feature Cross Enhancement (FCE) module that enhances query and exemplar features mutually to improve the accuracy of feature matching. Our approach outperforms state-of-the-art methods on multiple datasets, including FSC-147 and CARPK. Code is available at https://github.com/CBalance/SSD.

Yang Xu, Yifan Feng, Yue Gao

The limited availability of supervised labels (positive information) poses a notable challenge for open-world retrieval. However, negative information is more easily obtained but remains underexploited in current methods. In this paper, we introduce the Negative Prompt Driven Complementary Parallel Representation (NPCP) framework, which navigates the complexities of open-world retrieval through the lens of Negative Prompts. Specifically, we employ the Parallel Exclusive Embedding (PEE) to effectively utilize the prompt information, bilaterally capturing both explicit negative and implicit positive signals. To address the challenges of embedding unification and generalization, our method leverages high-order correlations among objects through the Complementary Structure Tuning (CST), by constructing a complementary hypergraph based on bi-directional and cross-category correlations. We have developed four multimodal datasets for open-world 3D object retrieval with negative prompts: NPMN, NPAB, NPNT, and NPES. Extensive experiments and ablation studies on these four benchmarks demonstrate the superiority of our method over current state-of-the-art approaches.

Kepeng Xu, Li Xu, Gang He, Wenxin Yu, Yunsong Li

Multiple complex degradations are coupled in low-quality video faces in the real world. Therefore, blind video face restoration is a highly challenging ill-posed problem, requiring not only hallucinating high-fidelity details but also enhancing temporal coherence across diverse pose variations. Restoring each frame independently in a naive manner inevitably introduces temporal incoherence and artifacts from pose changes and keypoint localization errors. To address this, we propose the first blind video face restoration approach with a novel parsing-guided temporal-coherent transformer (PGTFormer) without pre-alignment. PGTFormer leverages semantic parsing guidance to select optimal face priors for generating temporally coherent artifact-free results. Specifically, we pre-train a temporal-spatial vector quantized auto-encoder on high-quality video face datasets to extract expressive context-rich priors. Then, the temporal parse-guided codebook predictor (TPCP) restores faces in different poses based on face parsing context cues without performing face pre-alignment. This strategy reduces artifacts and mitigates jitter caused by cumulative errors from face pre-alignment. Finally, the temporal fidelity regulator (TFR) enhances fidelity through temporal feature interaction and improves video temporal consistency. Extensive experiments on face videos show that our method outperforms previous face restoration baselines. The code will be released on https://github.com/kepengxu/PGTFormer.

Guoxin Xiong, Yuan Wang, Zhaoyang Li, Wenfei Yang, Tianzhu Zhang, Xu Zhou, Shifeng Zhang, Yongdong Zhang

Point cloud few-shot semantic segmentation (PC-FSS) aims to segment objects within query samples of new categories given only a handful of annotated support samples. Although PC-FSS demonstrates enhanced category generalization capabilities compared to the fully supervised paradigm, the prevalent significant scene discrepancies, which can be systematically summarized into intra-semantic diversity and semantic inconsistency, have posed substantial challenges to the area. In this work, we design a novel Dual Enhancement Network (DENet) to comprehensively tackle different kinds of scene discrepancies in a coherent and synergistic framework. The proposed DENet enjoys several merits. First, we design a mutual aggregation module to reconcile the intrinsic tension between the support prototypes and query point features, and the intra-semantic diversity is diminished in a bidirectional manner. Second, the consistent purification strategy is introduced to eliminate ambiguous prototypes, thereby reducing the mismatches brought by semantic inconsistency. Extensive experiments on S3DIS and ScanNet under different settings demonstrate that DENet significantly outperforms previous SOTAs.

Jianqiang Xia, Dianxi Shi, Ke Song, Linna Song, Xiaolei Wang, Songchang Jin, Chenran Zhao, Yu Cheng, Lei Jin, Zheng Zhu 等

Most existing RGB-T tracking networks extract modality features in a separate manner, which lacks interaction and mutual guidance between modalities. This limits the network's ability to adapt to the diverse dual-modality appearances of targets and the dynamic relationships between the modalities. Additionally, the three-stage fusion tracking paradigm followed by these networks significantly restricts the tracking speed. To overcome these problems, we propose a unified single-stage Transformer RGB-T tracking network, namely USTrack, which unifies the above three stages into a single ViT (Vision Transformer) backbone through joint feature extraction, fusion and relation modeling. With this structure, the network can not only extract the fusion features of templates and search regions under the interaction of modalities, but also significantly improve tracking speed through the single-stage fusion tracking paradigm. Furthermore, we introduce a novel feature selection mechanism based on modality reliability to mitigate the influence of invalid modalities for final prediction. Extensive experiments on three mainstream RGB-T tracking benchmarks show that our method achieves the new state-of-the-art while achieving the fastest tracking speed of 84.2FPS. Code is available at https://github.com/xiajianqiang/USTrack.

Xing Xi, Yangyang Huang, Jinhao Lin, Ronghua Luo

Open-World Object Detection (OWOD) has garnered widespread attention due to its ability to recall unannotated objects. Existing works generate pseudo-labels for the model using heuristic priors, which limits the model’s performance. In this paper, we leverage the knowledge of the large-scale visual model to provide supervision for unknown categories. Specifically, we use the Segment Anything Model (SAM) to generate raw pseudo-labels for potential objects and refine them through Intersection over Union (IOU) and the shortest bounding box side length. Nevertheless, the abundance of pseudo-labels still exacerbates the competition issue in the one-to-many label assignment. To address this, we propose the Dual Matching Label Assignment (DMLA) strategy. Furthermore, we propose the Class-Awareness Neutralizer (CAN) to reduce the model’s bias towards known categories. Evaluation results on open-world object detection benchmarks, including MS COCO and Pascal VOC, show that our method achieves nearly 200% the unknown recall rate of previous state-of-the-art (SOTA) methods, reaching 41.5 U-Recall. Additionally, our approach does not add any extra parameters, maintaining the inference speed advantage of Faster R-CNN, leading the SOTA methods based on deformable DETR at a speed of over 10 FPS. Our code is available at https://github.com/xxyzll/KTCN.

Yong Wu, Yang Wang, Sanqing Qu, Zhijun Li, Guang Chen

We consider the problem of user-adaptive 3D gaze estimation. The performance of person-independent gaze estimation is limited due to interpersonal anatomical differences. Our goal is to provide a personalized gaze estimation model specifically adapted to a target user. Previous work on user-adaptive gaze estimation requires some labeled images of the target person data to fine-tune the model at test time. However, this can be unrealistic in real-world applications, since it is cumbersome for an end-user to provide labeled images. In addition, previous work requires the training data to have both gaze labels and person IDs. This data requirement makes it infeasible to use some of the available data. To tackle these challenges, this paper proposes a new problem called efficient label-free user adaptation in gaze estimation. Our model only needs a few unlabeled images of a target user for the model adaptation. During offline training, we have some labeled source data without person IDs and some unlabeled person-specific data. Our proposed method uses a meta-learning approach to learn how to adapt to a new user with only a few unlabeled images. Our key technical innovation is to use a generalization bound from domain adaptation to define the loss function in meta-learning, so that our method can effectively make use of both the labeled source data and the unlabeled person-specific data during training. Extensive experiments validate the effectiveness of our method on several challenging benchmarks.

Shengjian Wu, Li Sun, Qingli Li

DEtection TRansformer (DETR) becomes a dominant paradigm, mainly due to its common architecture with high accuracy and no post-processing. However, DETR suffers from unstable training dynamics. It consumes more data and epochs to converge compared with CNN-based detectors. This paper aims to stabilize DETR training through the online distillation. It utilizes a teacher model, accumulated by Exponential Moving Average (EMA), and distills its knowledge into the online model in following three aspects. First, the matching relation between object queries and ground truth (GT) boxes in the teacher is employed to guide the student, so queries within the student are not only assigned labels based on their own predictions, but also refer to the matching results from the teacher. Second, the teacher's initial query is given to the online student, and its prediction is directly constrained by the corresponding output from the teacher. Finally, the object queries from teacher's different decoding stages are used to build the auxiliary groups to accelerate the convergence. For each GT, two queries with the least matching costs are selected into this extra group, and they predict the GT box and participate the optimization. Extensive experiments show that the proposed OD-DETR successfully stabilizes the training, and significantly increases the performance without bringing in more parameters.

Jinmeng Wu, Chong Cao, Hao Zhang, Basura Fernando, Yanbin Hao, Hanyu Hong

The success of contrastive learning models like CLIP, known for aligning 2D image-text pairs, has inspired the development of triplet alignment for Large 3D Point Cloud Models (3D-PCM). Examples like ULIP integrate images, text, and point clouds into a unified semantic space. However, despite showing impressive zero-shot capabilities, frozen 3D-PCM still falls short compared to fine-tuned methods, especially when downstream 3D datasets are significantly different from upstream data. Addressing this, we propose a Data-Efficient, Training-Free 3D Adaptation method named PointTFA that adjusts ULIP outputs with representative samples. PointTFA comprises the Representative Memory Cache (RMC) for selecting a representative support set, Cloud Query Refactor (CQR) for reconstructing a query cloud using the support set, and Training-Free 3D Adapter (3D-TFA) for inferring query categories from the support set. A key advantage of PointTFA is that it introduces no extra training parameters, yet outperforms vanilla frozen ULIP, closely approaching few-shot fine-tuning training methods in downstream cloud classification tasks like ModelNet10 & 40 and ScanObjectNN. The code is available at: https://github.com/CaoChong-git/PointTFA.

Housheng Wei, Guanyu Xing, Jingwei Liao, Yanci Zhang, Yanli Liu

Video shadow detection faces significant challenges due to ambiguous semantics and variable shapes. Existing video shadow detection algorithms typically overlook the fine shadow details, resulting in inconsistent detection between consecutive frames in complex real-world video scenarios. To address this issue, we propose a spatial-temporal feature interaction strategy, which refines and enhances global shadow semantics with local prior features in the modeling of shadow relations between frames. Moreover, a structure-aware shadow prediction module is proposed, which focuses on modeling the distance relation between local shadow edges and regions. Quantitative experimental results demonstrate that our approach significantly outperforms the state-of-the-art methods, providing stable and consistent shadow detection results in complex video shadow scenarios.

Xiyu Wang, Anh-Dung Dinh, Daochang Liu, Chang Xu

Diffusion probabilistic models (DPMs) have been shown to generate high-quality images without the need for delicate adversarial training. The sampling process of DPMs is mathematically similar to Stochastic Gradient Descent (SGD), with both being iteratively updated with a function increment. Building on this, we present a novel reverse sampler for DPMs in this paper, drawing inspiration from the widely-used Adam optimizer. Our proposed sampler can be readily applied to a pre-trained diffusion model, utilizing momentum mechanisms and adaptive updating to enhance the generated image's quality. By effectively reusing update directions from early steps, our proposed sampler achieves a better balance between high-level semantics and low-level details. Additionally, this sampler is flexible and can be easily integrated into pre-trained DPMs regardless of the sampler used during training. Our experimental results on multiple benchmarks demonstrate that our proposed reverse sampler yields remarkable improvements over different baselines.

Wei Wang, Hanyang Li, Ke Shi, Chao Huang, Yang Cao, Cong Wang, Xiaochun Cao

Domain adaptation aims to achieve label transfer from a labeled source domain to an unlabeled target domain, where the two domains exhibit different distributions. Existing methods primarily concentrate on designing a feature extractor to learn better domain-invariant features, along with developing an effective classifier for reliable predictions. In this paper, we introduce optimal graph learning to generate a cross-domain graph that effectively connects the two domains, and two domain-specific graphs to capture domain-specific structures. On the one hand, we incorporate the three graphs into the label propagation (LP) classifier to enhance its robustness to distribution difference. On the other hand, we leverage the three graphs to introduce graph embedding losses, promoting the learning of locally discriminative and domain-invariant features. Furthermore, we maximize the nuclear norm of predictions in LP to enhance class diversity, thereby improving its robustness to class imbalance problem. Correspondingly, we develop an efficient algorithm to solve the associated optimization problem. Finally, we integrate the proposed LP and graph embedding losses into a deep neural network, resulting in our proposed deep cross-domain robust LP. Extensive experiments conducted on three cross-domain benchmark datasets demonstrate that our proposed approach could outperform existing state-of-the-art domain adaptation methods.

Song Wang, Jiawei Yu, Wentong Li, Hao Shi, Kailun Yang, Junbo Chen, Jianke Zhu

Semantic scene completion aims to infer the 3D geometric structures with semantic classes from camera or LiDAR, which provide essential occupancy information in autonomous driving. Prior endeavors concentrate on constructing the network or benchmark in a fully supervised manner. While the dense occupancy grids need point-wise semantic annotations, which incur expensive and tedious labeling costs. In this paper, we build a new label-efficient benchmark, named ScribbleSC, where the sparse scribble-based semantic labels are combined with dense geometric labels for semantic scene completion. In particular, we propose a simple yet effective approach called Scribble2Scene, which bridges the gap between the sparse scribble annotations and fully-supervision. Our method consists of geometric-aware auto-labelers construction and online model training with an offline-to-online distillation module to enhance the performance. Experiments on SemanticKITTI demonstrate that Scribble2Scene achieves competitive performance against the fully-supervised counterparts, showing 99% performance of the fully-supervised models with only 13.5% voxels labeled. Both annotations of ScribbleSC and our full implementation are available at https://github.com/songw-zju/Scribble2Scene.

Shuo Wang, Zhihao Wu, Jinwen Wang, Xiaobo Hu, Youfang Lin, Kai Lv

Despite the impressive success in visual control challenges, Visual Reinforcement Learning (VRL) policies have struggled to generalize to other scenarios. Existing works attempt to empirically improve the generalization capability, lacking theoretical support. In this work, we explore how to learn domain-invariant representations for VRL from an information-theoretical perspective. Specifically, we identify three Mutual Information (MI) terms. These terms highlight that a robust representation should preserve domain invariant information (return and dynamic transition) under significant observation perturbation. Furthermore, we relax the MI terms to derive three components for implementing a practical Mutual Information-based Invariant Representation (MIIR) algorithm for VRL. Extensive experiments demonstrate that MIIR achieves state-of-the-art generalization performance and the best sample efficiency in the DeepMind Control suite, Robotic Manipulation, and Carla.

Qi Wang, Yuchi Huo, Qi Ye, Rui Wang, Hujun Bao

Neural implicit surfaces with signed distance functions (SDFs) achieve superior quality in 3D geometry reconstruction. However, training SDFs is time-consuming because it requires a great number of samples to calculate accurate weight distributions and a considerable amount of samples sampled from the distribution for integrating the rendering results. Some existing sampling strategies focus on this problem. During the training, they assume a spatially-consistent convergence speed of kernel size, thus still suffering from low convergence or errors. Instead, we introduce an error-aware sampling method based on thin intervals of valid weight distributions, dubbed adaptive shells, to reduce the number of samples while still maintaining the reconstruction accuracy. To this end, we first extend Laplace-based neural implicit surfaces with learned spatially-varying kernel sizes which indicates the range of valid weight distributions. Then, the adaptive shell for each ray is determined by an efficient double-clipping strategy with spatially-varying SDF values and kernel sizes, fitting larger kernel sizes to wider shells. Finally, we calculate the error-bounded cumulative distribution functions (CDFs) of shells to conduct efficient importance sampling, achieving low-variance rendering with fewer calculations. Extensive results in various scenes demonstrate the superiority of our sampling technique, including significantly reducing sample counts and training time, even improving the reconstruction quality. The code is available at https://github.com/erernan/ESampling.

Cong Wang, Wei Wang, Chengjin Yu, Jie Mu

Patch-level non-local self-similarity is an important property of natural images. However, most existing methods do not consider this property into neural networks for image deraining, thus affecting recovery performance. Motivated by this property, we find that there exists significant patch recurrence property of a rainy image, that is, similar patches tend to recur many times in one image and its multi-scale images and external images. To better model this property for image detaining, we develop a multi-scale graph network with exemplars, called MSGNN, that contains two branches: 1) internal data-based supervised branch is used to model the internal relations of similar patches from the rainy image itself and its multi-scale images and 2) external data-participated unsupervised branch is used to model the external relations of the similar patches in the rainy image and exemplar. Specifically, we construct a graph model by searching the k-nearest neighboring patches from both the rainy images in a multi-scale framework and the exemplar. After obtaining the corresponding k neighboring patches from the multi-scale images and exemplar, we build a graph and aggregate them in an attentional manner so that the graph can provide more information from similar patches for image deraining. We embed the proposed graph in a deep neural network and train it in an end-to-end manner. Extensive experiments demonstrate that the proposed algorithm performs favorably against eight state-of-the-art methods on five public synthetic datasets and one real-world dataset. The source codes will be available at https://github.com/supersupercong/MSGNN.

Chenhui Wang, Tao Chen, Zhihao Chen, Zhizhong Huang, Taoran Jiang, Qi Wang, Hongming Shan

Despite their impressive generative performance, latent diffusion model-based virtual try-on (VTON) methods lack faithfulness to crucial details of the clothes, such as style, pattern, and text. To alleviate these issues caused by the diffusion stochastic nature and latent supervision, we propose a novel Faithful Latent Diffusion Model for VTON, termed FLDM-VTON. FLDM-VTON improves the conventional latent diffusion process in three major aspects. First, we propose incorporating warped clothes as both the starting point and local condition, supplying the model with faithful clothes priors. Second, we introduce a novel clothes flattening network to constrain generated try-on images, providing clothes-consistent faithful supervision. Third, we devise a clothes-posterior sampling for faithful inference, further enhancing the model performance over conventional clothes-agnostic Gaussian sampling. Extensive experimental results on the benchmark VITON-HD and Dress Code datasets demonstrate that our FLDM-VTON outperforms state-of-the-art baselines and is able to generate photo-realistic try-on images with faithful clothing details.

Bo Wang, Zhao Zhang, Mingbo Zhao, Xiaojie Jin, Mingliang Xu, Meng Wang

Mainstream image captioning models are usually two-stage captioners, i.e., encoding the region features by a pre-trained detector and then feeding them into a language model to generate the captions. However, such a two-stage procedure will lead to a task-based information gap that decreases the performance, because the region features in the detection task are suboptimal representations and cannot provide all the necessary information for subsequent captions generation. Besides, the region features are usually represented from the last layer of the detectors that lose the local details of images. In this paper, we propose a novel One-Stage Image Captioner (OSIC) with dynamic multi-sight learning, which directly transforms the images into descriptive sentences in one stage for eliminating the information gap. Specifically, to obtain rich features, multi-level features are captured by Swin Transformer, and then fed into a novel dynamic multi-sight embedding module to exploit both the global structure and local texture of input images. To enhance the global modeling capacity of the visual encoder, we propose a new dual-dimensional refining to non-locally model the features interaction. As a result, OSIC can directly obtain rich semantic information to improve the captioner. Extensive comparisons on the benchmark MS-COCO, Flickr8K and Flickr30K datasets verified the superior performance of our method.

Vivek Trivedy, Longin Jan Latecki

Common approaches to image retrieval include contrastive methods and specialized loss functions such as ranking losses and entropy regularizers. We present DMCAC (Divergence Minimization with Cross-Attention Classification), a novel image retrieval method that offers a new perspective on this training paradigm. We use self-supervision with a novel divergence loss framework alongside a simple data flow adjustment that minimizes a distribution over a database directly during training. We show that jointly learning a query representation over a database is a competitive and often improved alternative to traditional contrastive methods for image retrieval. We evaluate our method across several model configurations and four datasets, achieving state-of-the-art performance in multiple settings. We also conduct a thorough set of ablations that show the robustness of our method across full vs. approximate retrieval and different hyperparameter configurations.