Contrastive learning (CL) has emerged as a powerful paradigm for unsupervised feature representation, distinguishing between similar and dissimilar samples. However, traditional CL methods often struggle with intra-class variation, leading to insufficient robustness against noise and interference. In this paper, we propose a novel framework, Across Information Bottleneck-based Contrastive Learning (AIB-CL), which integrates the principles of the information bottleneck method to specifically address these variations. AIB-CL aims to find an optimal representation that retains maximal relevant information about the output while compressing irrelevant intra-class features. Through extensive experiments, we demonstrate that AIB-CL significantly enhances robustness to intra-class variance without sacrificing the discriminative power of feature representations. Our method shows advantages over existing contrastive learning models on standard benchmarks and exhibits superior performance in tasks sensitive to intra-class differences. The theoretical and empirical insights presented in this paper offer a new perspective on one of the fundamental challenges in unsupervised learning and provide a solid foundation for future research in feature representation learning.

We present CogniMap3D, a bioinspired framework for dynamic 3D scene understanding and reconstruction that emulates human cognitive processes. Our approach maintains a persistent memory bank of static scenes, enabling efficient spatial knowledge storage and rapid retrieval. CogniMap3D integrates three core capabilities: a multi-stage motion cue framework for identifying dynamic objects, a cognitive mapping system for storing, recalling, and updating static scenes across multiple visits, and a factor graph optimization strategy for refining camera poses. Given an image stream, our model identifies dynamic regions through motion cues with depth and camera pose priors, then matches static elements against its memory bank. When revisiting familiar locations, CogniMap3D retrieves stored scenes, relocates cameras, and updates memory with new observations. Evaluations on video depth estimation, camera pose reconstruction, and 3D mapping tasks demonstrate its state-of-the-art performance, while effectively supporting continuous scene understanding across extended sequences and multiple visits.

Current mainstream face recognition methods predominantly formulate the recognition task as a classification problem optimized via Margin Softmax loss. While this framework effectively facilitates the training of robust feature representation models—and has inspired a multitude of variants that steadily push the boundaries of recognition accuracy—it nonetheless suffers from inherent limitations when scaling to massive identity spaces. Specifically, such methods often grapple with training instability, a discrepancy between optimization objectives and real-world deployment demands, and the intricate challenge of hyperparameter tuning. To circumvent these obstacles, this paper introduces PartFace, a novel training paradigm for face recognition. Grounded in the principle of identity space partitioning, PartFace adopts a regression-based loss function formulated with Mean Squared Error. By strategically truncating the gradient contributions from non-neighbor categories, the model is steered to concentrate exclusively on the local similarity structures within the immediate neighborhood of a sample. This design markedly improves training stability and accelerates convergence. Comprehensive experimental evaluations across large-scale training datasets reveal that PartFace attains state-of-the-art or highly competitive performance on a wide array of standard benchmarks, such as LFW, CFP-FP, AgeDB, CALFW, CPLFW, IJB-B, and IJC-B. In direct comparison with leading contemporary approaches like TopoFR and AdaFace, PartFace distinguishes itself through more rapid initial convergence, significantly enhanced intra-class compactness, and greater inter-class separation margins, thereby substantiating its efficacy and preeminence in the domain of face recognition.

This paper focuses on research into multimodal pedestrian attribute recognition (PAR). Existing multimodal PAR methods face three major challenges: (1) reliance on fixed prompt templates that limit refined textual attribute descriptions, (2) neglect of essential visual details, such as textures and contours, which are crucial for distinguishing subtle attributes, and (3) insufficient interaction between visual and textual modalities, resulting in suboptimal recognition performance. To address these challenges, we propose a Frequency Enhancement and Mamba-Driven Representation Learning (FreqMamba) framework. Specifically, we design a Dynamic Prompt Strategy (DPS) that dynamically adapts textual descriptions based on visual features, resulting in refined textual descriptions. Additionally, a Frequency-Enhanced Module (FEM) is introduced to capture texture and contour details in the frequency domain, enriching visual features and further enhancing the generation of dynamic prompts. Finally, we propose a Mamba-Driven Modality Interaction (MDMI) method combined with a Bidirectional Modality Contrastive Loss (BMCLoss) to facilitate effective interaction between modalities, integrating visual and textual information comprehensively. Extensive experiments on multiple PAR datasets, including PETA, PA100K, RAPv1, and RAPv2, demonstrate the superiority of our FreqMamba approach over existing methods.

Pedestrian Attribute Recognition (PAR) plays a crucial role in various computer vision applications, demanding precise and reliable identification of attributes from pedestrian images. Traditional PAR methods, though effective in leveraging attention mechanisms, often suffer from the lack of direct supervision on attention, leading to potential overfitting and misallocation. This paper introduces a novel and model-agnostic approach, Attention-Aware Regularization (AAR), which rethinks the attention mechanism by integrating causal reasoning to provide direct supervision of attention maps. AAR employs perturbation techniques and a unique optimization objective to assess and refine attention quality, encouraging the model to prioritize attribute-specific regions. Our method demonstrates significant improvement in PAR performance by mitigating the effects of incorrect attention and fostering a more effective attention mechanism. Experiments on standard datasets showcase the superiority of our approach over existing methods, setting a new benchmark for attention-driven PAR models.

The main purpose of domain adaptation (DA) is to conduct cross-domain related knowledge transfer. Considering the issue of unsupervised DA (UDA), learning a transformation that reduces the differences between domains is the primary goal. In addition to minimizing both the marginal and conditional distributions between the source and target domains, many methods explore potential factors that show the commonalities among domains to yield improved learning efficiency. However, geometric structure information is overlooked by most existing approaches, indicating that the shared information between domains has not been fully exploited. On account of this finding, by taking advantage of more potential shared factors to further enhance the results of DA, we propose an approximate geometric structure transfer (AGST) method for cross-domain image classification in this paper. By combining structural consistency and sample reweighting techniques, AGST encodes the geometric structure information taken from the samples in both domains, enabling it to easily obtain richer interdomain features and effectively facilitate knowledge transfer. Extensive experiments are conducted on several cross-domain data benchmarks. The experimental results indicate that our AGST method can outperform many state-of-the-art algorithms.

Person re-identification (Re-ID) plays a crucial role in the domains of security surveillance and pedestrian behavior analysis, as it aims to retrieve specific individuals captured by different cameras. However, the task of Re-ID remains immensely challenging in the field of computer vision, primarily due to the extensive intra-class variations exhibited by individuals across cameras. These variations include occlusions, illuminations, viewpoints, and poses. In this paper, we present a novel Re-ID framework that addresses the inherent issues related to intra-class variations. Our proposed approach incorporates both auxiliary-domain classification (ADC) and layered semi-second-order information bottleneck (LyrS2IB) techniques. By incorporating ADC as an auxiliary task, we leverage coarse-grained essential features that effectively distinguish individuals from the background. This enables the development of both coarse- and fine-grained feature representations for Re-ID. Furthermore, our framework integrates LyrS2IB to handle redundancy, irrelevance, and noise present in Re-ID features resulting from intra-class variations. This integration allows us to compress and optimize these features without incurring additional computation overhead during inference. Extensive experiments validate the efficacy of our proposed method, demonstrating a significant reduction in the neural network output variance of intra-class person images, firmly establishing the superior performance of our approach in the field of Re-ID.

We present Motion Marionette, a zero-shot framework for rigid motion transfer from monocular source videos to single-view target images. Previous works typically employ geometric, generative, or simulation priors to guide the transfer process, but these external priors introduce auxiliary constraints that lead to trade-offs between generalizability and temporal consistency. To address these limitations, we propose guiding the motion transfer process through an internal prior that exclusively captures the spatial-temporal transformations and is shared between the source video and any transferred target video. Specifically, we first lift both the source video and the target image into a unified 3D representation space. Motion trajectories are then extracted from the source video to construct a spatial-temporal (SpaT) prior that is independent of object geometry and semantics, encoding relative spatial variations over time. This prior is further integrated with the target object to synthesize a controllable velocity field, which is subsequently refined using Position-Based Dynamics to mitigate artifacts and enhance visual coherence. The resulting velocity field can be flexibly employed for efficient video production. Empirical results demonstrate that Motion Marionette generalizes across diverse objects, produces temporally consistent videos that align well with the source motion, and supports controllable video generation.

Multi-label Pedestrian Attribute Recognition (PAR) involves identifying a series of semantic attributes in person images. Existing PAR solutions typically rely on CNN as the backbone network to extract pedestrian features. Unfortunately, CNNs process only one adjacent region at a time, resulting in the disappearance of long-range relations between different attribute-specific regions. To address this limitation, we adopt the Vision Transformer (ViT) instead of CNN as the backbone for PAR, aiming to build long-range relations and extract more robust features. However, PAR suffers from an inherent attribute imbalance issue, causing ViT to naturally focus more on attributes that appear frequently in the training set and ignore some pedestrian attributes that appear less. The native features extracted by ViT are not able to tolerate the imbalance attribute distribution issue. To tackle this issue, we propose a novel component and a dual-level loss: the Selective Feature Activation Method (SFAM), the Orthogonal Feature Activation Loss (OFALoss), and Orthogonal Weight Regularization Loss (OWRLoss). SFAM smartly suppresses the more informative attribute-specific features, thus compelling the PAR model to pay greater attention to attribute-specific regions that are often overlooked. The proposed OFALoss enforces an orthogonal constraint on the original feature extracted by ViT and the suppressed features from SFAM, promoting the comprehensiveness of feature representation in each attribute-specific region. Furthermore, OWRLoss is employed for decreasing correlations among entries of the last shared classification layer, which can alleviate the highly correlated of weight vectors caused by non-uniform distribution. This can prevent excessive mutual interference among different attributes during attribute recognition. Our model-agnostic approach is plug-and-play, requiring no additional training parameters in the training process. We conduct experiments on several benchmark PAR datasets, including PETA, PA100K, RAPv1, and RAPv2, demonstrating the effectiveness of our method. Specifically, our method outperforms existing state-of-the-art approaches.

Dataset distillation aims to synthesize a small dataset from a large dataset, enabling the model trained on it to perform well on the original dataset. With the blooming of large language models and multimodal large language models, the importance of multimodal datasets, particularly image-text datasets, has grown significantly. However, existing multimodal dataset distillation methods are constrained by the Matching Training Trajectories algorithm, which significantly increases the computing resource requirement, and takes days to process the distillation. In this work, we introduce EDGE, a generative distillation method for efficient multimodal dataset distillation. Specifically, we identify two key challenges of distilling multimodal datasets with generative models: 1) The lack of correlation between generated images and captions. 2) The lack of diversity among generated samples. To address the aforementioned issues, we propose a novel generative model training workflow with a bi-directional contrastive loss and a diversity loss. Furthermore, we propose a caption synthesis strategy to further improve text-to-image retrieval performance by introducing more text information. Our method is evaluated on Flickr30K, COCO, and CC3M datasets, demonstrating superior performance and efficiency compared to existing approaches. Notably, our method achieves results 18$\times$ faster than the state-of-the-art method. Our code will be made public at https://github.com/ichbill/EDGE.

Pedestrian attribute recognition (PAR) seeks to predict multiple semantic attributes associated with a specific pedestrian. There are two types of approaches for PAR: unimodal framework and bimodal framework. The former one is to seek a robust visual feature. However, the lack of exploiting semantic feature of linguistic modality is the main concern. The latter one utilizes prompt learning techniques to integrate linguistic data. However, static prompt templates and simple bimodal concatenation cannot to capture the extensive intra-class attribute variability and support active modalities collaboration. In this paper, we propose an Enhanced Visual-Semantic Interaction with Tailored Prompts (EVSITP) framework for PAR. We present an Image-Conditional Dual-Prompt Initialization Module (IDIM) to adaptively generate context-sensitive prompts from visual inputs. Subsequently, a Prompt Enhanced and Regularization Module (PERM) is proposed to strengthen linguistic information from IDIM. We further design a Bimodal Mutual Interaction Module (BMIM) to ensure bidirectional modalities communication. In addition, existing PAR datasets are collected over a short period in limited scenarios, which do not align with real-world scenarios. Therefore, we annotate a long-term person re-identification dataset to create a new PAR dataset, Celeb-PAR. Experiments on several challenging PAR datasets show that our method outperforms state-of-the-art approaches.

X-ray imaging is indispensable in medical diagnostics, yet its use is tightly regulated due to potential health risks. To mitigate radiation exposure, recent research focuses on generating novel views from sparse inputs and reconstructing Computed Tomography (CT) volumes, borrowing representations from the 3D reconstruction area. However, these representations originally target visible light imaging that emphasizes reflection and scattering effects, while neglecting penetration and attenuation properties of X-ray imaging. In this paper, we introduce X-Field, the first 3D representation specifically designed for X-ray imaging, rooted in the energy absorption rates across different materials. To accurately model diverse materials within internal structures, we employ 3D ellipsoids with distinct attenuation coefficients. To estimate each material's energy absorption of X-rays, we devise an efficient path partitioning algorithm accounting for complex ellipsoid intersections. We further propose hybrid progressive initialization to refine the geometric accuracy of X-Filed and incorporate material-based optimization to enhance model fitting along material boundaries. Experiments show that X-Field achieves superior visual fidelity on both real-world human organ and synthetic object datasets, outperforming state-of-the-art methods in X-ray Novel View Synthesis and CT Reconstruction.

Pedestrian attribute recognition (PAR) involves accurately identifying multiple attributes present in pedestrian images. There are two main approaches for PAR: part-based method and attention-based method. The former relies on existing segmentation or region detection methods to localize body parts and learn corresponding attribute-specific feature from the corresponding regions, where the performance heavily depends on the accuracy of body region localization. The latter adopts the embedded attention modules or transformer attention to exploit detailed feature. However, it can focus on certain body regions but often provide coarse attention, failing to capture fine-grained details, the learned feature may also be interfered with by irrelevant information. Meanwhile, these methods overlook the global contextual information. This work argues for replacing coarse attention with detailed attention and integrating it with global contextual feature from ViT to jointly represent attribute-specific regions. To tackle this issue, we propose a High-order Diversity Feature Learning (HDFL) method for PAR based on ViT. We utilize a polynomial predictor to design an Attribute-specific Detailed Feature Exploration (ADFE) module, which can construct the high-order statistics and gain more fine-grained feature. Our ADFE module is a parameter-friendly method that provides flexibility in deciding its utilization during the inference phase. A Soft-redundancy Perception Loss (SPLoss) is proposed to adaptively measure the redundancy between feature of different orders, which can promote diverse characterization of features. Experiments on several PAR datasets show that our method achieves a new state-of-the-art (SOTA) performance. On the most challenging PA100K dataset, our method outperforms previous SOTA by 1.69% and achieves the highest mA of 84.92%.

The recent introduction of diffusion models in dataset distillation has shown promising potential in creating compact surrogate datasets for large, high-resolution target datasets, offering improved efficiency and performance over traditional bi-level/uni-level optimization methods. However, current diffusion-based dataset distillation approaches overlook the evaluation process and exhibit two critical inconsistencies in the distillation process: (1) Objective Inconsistency, where the distillation process diverges from the evaluation objective, and (2) Condition Inconsistency, leading to mismatches between generated images and their corresponding conditions. To resolve these issues, we introduce Condition-aware Optimization with Objective-guided Sampling (CaO_2), a two-stage diffusion-based framework that aligns the distillation process with the evaluation objective. The first stage employs a probability-informed sample selection pipeline, while the second stage refines the corresponding latent representations to improve conditional likelihood. CaO_2 achieves state-of-the-art performance on ImageNet and its subsets, surpassing the best-performing baselines by an average of 2.3

Deep learning has made remarkable progress recently, largely due to the availability of large, well-labeled datasets. However, the training on such datasets elevates costs and computational demands. To address this, various techniques like coreset selection, dataset distillation, and dataset quantization have been explored in the literature. Unlike traditional techniques that depend on uniform sample distributions across different classes, our research demonstrates that maintaining performance is feasible even with uneven distributions. We find that for certain classes, the variation in sample quantity has a minimal impact on performance. Inspired by this observation, an intuitive idea is to reduce the number of samples for stable classes and increase the number of samples for sensitive classes to achieve a better performance with the same sampling ratio. Then the question arises: how can we adaptively select samples from a dataset to achieve optimal performance? In this paper, we propose a novel active learning based adaptive sampling strategy, Dataset Quantization with Active Learning based Adaptive Sampling (DQAS), to optimize the sample selection. In addition, we introduce a novel pipeline for dataset quantization, utilizing feature space from the final stage of dataset quantization to generate more precise dataset bins. Our comprehensive evaluations on the multiple datasets show that our approach outperforms the state-of-the-art dataset compression methods.

Pedestrian Attribute Recognition (PAR) involves identifying the attributes of individuals in person images. Existing PAR methods typically rely on CNNs as the backbone network to extract pedestrian features. However, CNNs process only one adjacent region at a time, leading to the loss of long-range inter-relations between different attribute-specific regions. To address this limitation, we leverage the Vision Transformer (ViT) instead of CNNs as the backbone for PAR, aiming to model long-range relations and extract more robust features. However, PAR suffers from an inherent attribute imbalance issue, causing ViT to naturally focus more on attributes that appear frequently in the training set and ignore some pedestrian attributes that appear less. The native features extracted by ViT are not able to tolerate the imbalance attribute distribution issue. To tackle this issue, we propose two novel components: the Selective Feature Activation Method (SFAM) and the Orthogonal Feature Activation Loss. SFAM smartly suppresses the more informative attribute-specific features, compelling the PAR model to capture discriminative features from regions that are easily overlooked. The proposed loss enforces an orthogonal constraint on the original feature extracted by ViT and the suppressed features from SFAM, promoting the complementarity of features in space. We conduct experiments on several benchmark PAR datasets, including PETA, PA100K, RAPv1, and RAPv2, demonstrating the effectiveness of our method. Specifically, our method outperforms existing state-of-the-art approaches by GRL, IAA-Caps, ALM, and SSC in terms of mA on the four datasets, respectively.

Long-term (also called Clothing-Change) person re-identification (CC-reID) aims at confirming the identity of pedestrians captured at diverse locations and/or times. Current CC-reID methods heavily rely on ID features learned by the CNN architecture. However, with limited receptive fields, CNN is hard to effectively explore some unique but discriminative ID features (e.g., hair style, tattoo and accessories) from small body regions. Compared with CNN, Transformer has certain merits in exploring more diverse ID-unique features 1 and retaining more details by the multi-head self-attention design and the removal of down-sampling operation. In this paper, a two-stream hybrid Convolution-Transformer Network (CT-Net) is proposed for CC-reID by combining both CNN and Transformer parallelly in an end-to-end learning scheme. Specifically, CT-Net contains a CNN-based stream (C-Stream) and a Transformer-based stream (T-Stream). Compared with using C-Stream only, T-Stream is used to encourage the C-Stream to explore more detailed ID-unique features when the clothing information is no reliable in CC-reID. Specifically, a Feature Supplement Module (FSM) is proposed to transfer features learned by T-Stream to C-Stream from low-level to high-level for mining more ID-unique feature. In order to further enhance the discriminability 2 and complementary of ID features learned by our CT-Net, we also introduce a hierarchical supervision with bilinear pooling (HSBP). Experimental results demonstrate that CT-Net performs favorably against the state-of-the-art methods over three CC-reID benchmarks. Meanwhile, CT-Net also demonstrates good generalization ability by achieving comparable performance on traditional person re-ID datasets such as Market-1501 and DukeMTMC-reID.

Cross-modality face recognition aims to match facial images across different modalities. This task becomes very challenging when one of the modalities is the facial caricature, which enhances instinctive facial features through extreme distortions and exaggerations with diverse styles by artists. In this paper, we develop a novel modality interference decoupling and representation alignment (MIR) method for visual-caricature face recognition. Our MIR method consists of a backbone network, an identity-interference orthogonal decoupling (IIOD) module, and a modality feature alignment (MFA) module. The IIOD module adopts a three-branch structure to decouple the deep semantic features extracted by the backbone network into identity features and modality features. In IIOD, we design an identity subspace alignment (ISA) module to align the identity features from different branches. Moreover, we design the MFA module to perform feature alignment between the modality feature from the IIOD module and that from the pre-trained modality interference information encoder (MIE) via adversarial learning, extracting the modality-specific information. Based on the above designs, we can effectively alleviate the interference of modality differences and style differences, improving the final performance. Extensive experimental results on multiple datasets show that our proposed method outperforms several state-of-the-art cross-modality face recognition methods.

Multi-label pedestrian attribute recognition (PAR) involves assigning multiple attributes to pedestrian images captured by video surveillance cameras. Despite its importance, learning robust attribute-related features for PAR remains a challenge due to the large intra-attribute variations in the image space. These variations, which stem from changes in pedestrian poses, illumination conditions, and background noise, make extracted attribute-related features susceptible to irrelevant information or noise interference. Existing PAR methods rely on body prior extractors or attention mechanisms to locate attribute-correlation regions for extracting robust features. However, these methods may not be robust to intra-attribute variations, which limits their effectiveness. To address this challenge, we propose a novel and flexible PAR framework that leverages the exponential information bottleneck (ExpIB) approach. Our ExpIB-Net uses mutual information compression as the main penalty during the early stage of training, thereby eliminating irrelevant information. As training progresses, the mutual information penalty weakens and the Binary Cross-Entropy Loss (BCELoss) contributes to improving the PAR recognition accuracy. Our method can also be integrated into an attention module to form the AttExpIB-Net, which better handles intra-attribute variations for better performance. Additionally, our model-agnostic ExpIB approach is plug-and-play, requiring no additional computational overhead during inference. Experiments on several challenging PAR datasets show that our method outperforms state-of-the-art approaches.

Spiking neural networks (SNNs) have captivated the attention worldwide owing to their compelling advantages in low power consumption, high biological plausibility, and strong robustness. However, the intrinsic latency associated with SNNs during inference poses a significant challenge, impeding their further development and application. This latency is caused by the need for spiking neurons to collect electrical stimuli and generate spikes only when their membrane potential exceeds a firing threshold. Considering the firing threshold plays a crucial role in SNN performance, this article proposes a self-driven adaptive threshold plasticity (SATP) mechanism, wherein neurons autonomously adjust the firing thresholds based on their individual state information using unsupervised learning rules, of which the adjustment is triggered by their own firing events. SATP is based on the principle of maximizing the information contained in the output spike rate distribution of each neuron. This article derives the mathematical expression of SATP and provides extensive experimental results, demonstrating that SATP effectively reduces SNN inference latency, further reduces the computation density while improving computational accuracy, so that SATP facilitates SNN models to be with low latency, sparse computing, and high accuracy.