Weakly supervised semantic segmentation (WSSS) provides a compelling solution to reduce annotation costs in autonomous driving perception systems. However, existing methods fail to meet the distinct challenges of driving scenes, such as dramatic scale variations, intricate spatial relationships, and stringent boundary precision demands. To address these, we introduce DyFuse, a novel framework that seamlessly blends CLIP’s semantic understanding with SAM’s robust visual representations for WSSS in driving scenes. Our approach introduces three core innovations: (1) a Multi-scale Context Perception Enhancement (MCPE) module that captures objects across scales via parallel processing branches; (2) a dynamic fusion mechanism that adaptively merges complementary features based on reliability estimates; and (3) an intelligent pseudo-label strategy that combines CLIP’s semantic insights with SAM’s boundary accuracy. Extensive experiments demonstrate that DyFuse outperforms previous methods, increasing mIoU by 23.3% on Cityscapes, 26.5% on CamVid, and 18.3% on WildDash2, with consistent robustness in adverse weather. Remarkably, using only image-level labels, it narrows the gap with fully supervised methods to under 3% in Cityscapes and even surpasses them on CamVid validation set. Ablation studies highlight the impact of each component, especially in enhancing the segmentation of small objects and the complex boundary delineation that is critical to driving safety.

Graph Convolutional Networks (GCNs) have become a cornerstone in skeleton-based action recognition due to their ability to model non-Euclidean body structures. However, their capacity to construct expressive and structured representations is fundamentally constrained by the tight coupling of static graph topology represented by adjacency matrix. To address this, we introduce a novel architectural principle called Hierarchical Kernel Decoupling (HKD), which disentangles spatial aggregation semantics from fixed topological constraints, thereby enabling more structured and flexible representation learning. Our approach introduces two progressive modules. The first module initially decouples the adjacency matrix using a K-hop neighborhood partitioning strategy to build a multi-level convolutional neighborhood, allowing the model to learn from both local and distant joints in a structured and scalable manner. This partitioning strategy expands the receptive field while maintaining topological coherence, enabling diverse feature extraction. The second module further decouples the multiple joints within same partition by adaptively adjusting the contribution of each node within its partition using a context-aware partition enhancer. Together, these modules implement a hierarchical decoupling mechanism that transforms fixed, topology-coupled kernels into adaptive, structure-aware representations suited for skeleton-based action recognition. Integrated with a multi-branch temporal modeling design, the proposed network achieves state-of-the-art performance on three large-scale benchmarks: NTU RGB+D, NTU RGB+D 120, and NW-UCLA. Extensive experiments demonstrate that HKD-Graph Convolutional Network (HKD-GCN) not only improves accuracy but also perceives complex human actions with computational efficiency.

Scene Graph Generation (SGG) aims to parse images into structured graph representations containing objects and their interrelationships, serving as a critical task for achieving advanced scene understanding. However, current SGG techniques exhibit significant bias due to the long-tailed distribution of predicates, with models tending to predict head predicates while neglecting tail predicates. While existing debiasing methods effectively address statistical imbalance through data or loss adjustments, they primarily operate at the distribution level, lacking fine-grained discriminative capability for semantically similar predicates and failing to exploit the synergistic effects of triplets. To address these limitations, this paper proposes the ECORE-SGG framework, which achieves unbiased SGG through evidence-aware representation learning and coherent relational modeling. The framework progressively optimizes relationship prediction through three synergistic core modules. First, the Multi-modal Evidence-aware Fusion (MEF) module integrates visual and semantic evidence from union regions to enhance discriminative capability for semantically similar predicates. Building upon this, the Context-Aware Triplet modeling (CAT) module constructs dual-channel visual-semantic representations to model inter-triplet dependencies and fuse multi-modal contextual cues, ensuring scene graph global consistency. Furthermore, the Adaptive Uncertainty Modeling (AUM) module quantifies aleatoric uncertainty and incorporates it into training via adaptive sample weighting, improving robustness on ambiguous samples. Experimental results on the Visual Genome, GQA, and Open Images V6 datasets demonstrate that the proposed method outperforms existing state-of-the-art approaches, with ablation analysis further confirming the effectiveness of the ECORE mechanism. The code is available at https://github.com/VsGraph/ECORE.

Abstract Scene graph generation (SGG) aims to detect objects and recognize their semantic relationships from images, serving as an important bridge between visual perception and high-level semantic understanding. Existing debiasing methods primarily address inter-class long-tailed bias, while paying insufficient attention to intra-class semantic heterogeneity of predicates. High-frequency predicates typically encompass multiple semantic sub-patterns, causing instances of the same predicate to scatter widely in the visual feature space and resulting in a fundamental misalignment between feature distributions and classification objectives. To address these issues, we propose AMCA, an unbiased SGG framework integrating adaptive multi-prototype learning with cross-modal alignment. Specifically, the Multi-Prototype Construction (MPC) module employs a distribution-aware prototype allocation strategy to quantify the semantic distributional complexity of each predicate and dynamically assign prototype counts, explicitly modeling intra-class semantic heterogeneity. Building on this, the Hybrid Relation Refinement Module (HRRM) embeds semantic prototypes into the visual relation feature refinement process via a prototype-augmented attention mechanism, and further employs Mamba to model global contextual dependencies among different relation representations, achieving deep fusion of semantic priors and visual features. Furthermore, the Visual-Text Alignment (VTA) module aligns the fused visual relation representations with CLIP text embeddings, significantly enhancing the model’s ability to distinguish fine-grained semantics. Extensive experiments on Visual Genome and GQA demonstrate that AMCA achieves consistently competitive performance across multiple SGG tasks, with particularly strong improvements on the unbiased mR@K metric. Our code is publicly available at https://github.com/SJVision/AMCA .

Gait recognition, as a critical biometric technology, continues to face challenges under existing paradigms, including error accumulation from upstream tasks and insufficient semantic understanding. Moreover, real-world deployment demands the capability to continuously learn new identities and variations without succumbing to catastrophic forgetting. To address these issues, we propose SemGait, a novel multimodal lifelong gait recognition framework that integrates structured semantic descriptions with visual information. Specifically, we introduce the Gait Semantic Description Framework (GSDF), which converts low-level kinematic features from 2D skeleton sequences into rich, interpretable natural language descriptions, thereby bridging the gap in textual semantics within gait analysis. To harness the complementary advantages of multimodal inputs, SemGait adopts a multi-stage feature learning pipeline comprising Specific and Shared Feature Decoupling (SSFD) to disentangle modality-specific and shared representations; Gait Attribute Knowledge Encoder (GAKE) to capture fine-grained attribute-level semantics; and Attribute-Driven Mixture of Experts (ADMoE) to enable discriminative fusion of visual, textual, and attribute features. To facilitate effective lifelong learning, we further propose Lifelong Gait Learning Strategies (LGLS), which include the Attribute-Anchored Forgetting Suppression (AFS) mechanism for preserving prior knowledge and the Adaptive Knowledge Transfer (AKT) mechanism for enhancing generalization to new domains. Extensive experiments on four public benchmark datasets demonstrate that SemGait consistently surpasses state-of-the-art methods in both within-domain and cross-domain gait recognition. Furthermore, in lifelong learning scenarios, it exhibits superior robustness against catastrophic forgetting while maintaining strong adaptability to new conditions, highlighting the effectiveness of the proposed multimodal and lifelong learning strategies.

Parsing-based gait recognition has emerged as a promising paradigm for handling complex in-the-wild scenarios by leveraging fine-grained semantic body parts. However, existing methods typically treat all spatial regions equally, neglecting the inherent boundary ambiguity caused by low resolution, motion blur, and imperfect parsing predictions. To address these challenges, we propose a framework termed Robust Boundary-aware Parsing Gait (RBPGait). First, we devise a Boundary-aware Semantic Refinement (BASR) module to explicitly model spatial uncertainty. By estimating a pixel-wise boundary uncertainty map, the BASR module dynamically recalibrates the interaction between global context and local details via a dual-stream modulation mechanism, ensuring that ambiguous boundary regions are effectively refined. Second, to mitigate the sensitivity to outliers and label noise, we introduce a Hybrid Hard-aware Triplet Loss (HHTL). Unlike conventional hard mining strategies, this loss stabilizes optimization by smoothing the gradients derived from the top-k hardest sample pairs. Extensive experiments on two challenging datasets demonstrate that RBPGait achieves state-of-the-art performance. Specifically, on the Gait3D dataset, our method outperforms the leading silhouette-based approach by 11.6% in Rank-1 accuracy and 13.0% in mAP, while surpassing the strong parsing-based baseline by 8.8% in mINP. Furthermore, on the clothing-varied CCGR-Mini dataset, it achieves significant improvements of 6.2% and 9.7% in Rank-1 and mINP, respectively, verifying its superior robustness in unconstrained environments.

Weakly supervised semantic segmentation (WSSS) offers a promising solution to reduce annotation costs in autonomous driving perception systems. However, existing methods struggle with the complex environmental conditions inherent to real-world driving scenarios, including adverse weather, variable lighting, and challenging visibility conditions. To address these limitations, we introduce EASeg, a novel framework that enhances segmentation robustness across diverse environmental conditions while requiring only image-level supervision. Our approach introduces three key innovations: (1) a multi-scale feature module that captures objects at varying scales followed by a boundary-aware enhancement component for precise delineation; (2) a dual-stream environmental adaptation mechanism that separately models global weather patterns and local illumination variations; and (3) a reliability-guided feature integration strategy that dynamically combines backbone features with foundation models based on their estimated reliability. Extensive experiments demonstrate that EASeg outperforms previous best methods, increasing mIoU by 24.5% on Cityscapes, 27.5% on CamVid, and 22.5% on WildDash2. Ablation studies confirm that our work represents a significant advancement toward practical, all-weather autonomous driving systems that enhance safety through improved segmentation of small objects and precise boundary delineation, while minimizing annotation requirements.

With the evolution of cloud-native microservice architectures traditional sidecar-based monitoring patterns and fragmented security tools have introduced significant resource overhead and management complexity. Current research indicates that the sidecar pattern induces notable network latency in distributed tracing while hybrid deployments of multiple kernel-level security tools often lead to policy execution conflicts and performance bottlenecks. Furthermore native probing mechanisms in current container orchestration platforms lack precise awareness of internal application states and malware detection methods based on simple system call counting struggle to identify highly concealed advanced persistent threats. To address these limitations this paper proposes an adaptive and unified framework for full-stack observability and security driven by extended Berkeley Packet Filter technology. This framework innovatively constructs a kernel-level closed loop for monitoring analysis and response. First a multi-modal data fusion adaptive sensing mechanism is proposed to unify the collection of application-layer network flows and system call sequences within the kernel space. This approach eliminates the need for sidecars while providing context-rich feature data for anomaly detection models. Second a state-aware dynamic hook orchestration algorithm is designed. Operating in a low-load monitoring mode by default this algorithm automatically escalates hook depth to the Linux Security Module layer for active blocking or granular forensics upon detecting service anomalies or behavioral drift effectively solving the performance dilemma of continuous high-load monitoring. Finally a graph-theory-driven unified policy arbitration engine is constructed to pre-calculate and resolve logical conflicts between security defense policies and operational scheduling policies prior to bytecode loading. Experimental results demonstrate that the proposed framework integrates full-stack observability and real-time security defense within a single agent significantly reducing end-to-end latency for anomaly detection and response while drastically lowering CPU overhead effectively defending against complex supply chain attacks and resource exhaustion attacks while maintaining a low false-positive rate.

In response to the issue of the GaitSet algorithm's backbone network being ineffective in feature learning and easily ignoring fine-grained features in local areas, we propose a gait recognition algorithm based on multi-scale fine-grained feature fusion (MSFG-Gaitset). This algorithm reconstructs the GaitSet network by introducing a multi-scale feature extraction module and a local feature extraction module, which allows for the extraction of multi-scale features while dividing the feature map into multiple parts for fine-grained feature extraction. Additionally, we improve the hard-batch triplet loss function to optimize the gait model, making the training process more effective. To further enhance the model’s performance, features at different layers are fused to obtain richer feature representations. To validate the algorithm’s effectiveness, we tested it on the CASIA-B dataset, and the results showed that the gait recognition accuracy reached 85.81%, a 2.6% improvement over GaitSet. This demonstrates that the MSFG-GaitSet algorithm can effectively mitigate the negative impact of complex environments on gait recognition performance and achieve high-accuracy gait recognition in complex scenarios. Experimental data fully demonstrate that this method not only has high gait recognition accuracy but also exhibits good robustness and wide applicability.

Objective: Text-Based Person Retrieval (TBPR) aims to match target images using natural language descriptions, yet it faces significant challenges such as complex visual structures, diverse semantic expressions, and limited annotated data. These issues lead to intra-modal knowledge uncertainty and weak inter-modal correlations. Existing approaches predominantly focus on explicit alignment between heterogeneous modalities, often overlooking the latent associations within homogeneous knowledge. This study addresses these limitations to enhance both the performance and efficiency of TBPR. Method: This paper proposes a Multivariate Enhancement Fine-Grained Knowledge Progressive Network (ME-FKPN) to address the challenge of text-based person retrieval across modalities. ME-FKPN enhances the synergy between homogeneous and heterogeneous knowledge in a hierarchical manner, progressively establishing more accurate semantic alignments between images and texts. The framework comprises three key innovations: Standardized Knowledge Anchor (SKA) constructs a knowledge graph to standardize semantics; Mixed of LoRA Experts (MoLE) integrates defogged color and grayscale images to extract multi-level visual features; Multivariate Knowledge Progressive Optimization Strategy (MKPOS) achieves steady performance improvements through hierarchical augmentation and staged training. Novelty: The proposed ME-FKPN model introduces a novel solution to the TBPR task by integrating hierarchical knowledge enhancement and progressive optimization strategies. By deeply mining latent associations among homogeneous knowledge and fostering collaborative representation of heterogeneous knowledge, the approach effectively overcomes the limitations of existing techniques in handling complex cross-modal relationships and data sparsity. Findings: ME-FKPN outperforms all state-of-the-art TBPR models across three public datasets. Notably, on the challenging UFine6926 ultra-fine-grained dataset, our method achieves improvements of 16.49%, 9.79%, 6.23%, and 14.69% on R@1, R@5, R@10, and mAP metrics, respectively, compared to mainstream approaches.

This study advances the utilization of semantic information in person re-identification (ReID) by leveraging pre-trained vision-language models, addressing the current limitations in semantic information processing within ReID systems. While recent studies have explored CLIP integration for ReID tasks, their training approaches have inadvertently diminished semantic information by focusing primarily on indirect alignment between person IDs through text encoders and image features. Through comprehensive empirical analysis of semantic information’s role in pedestrian ReID, we propose MoSCE-ReID, a mixed semantic clustering expert model. The framework incorporates two key components: a learnable Attribute Group Weight Extractor (AGWE) and a Mixed of LoRA Expert (MoLE) module, designed specifically for attribute group feature extraction. The final ReID decisions are made through the synergistic integration of attribute group features and global features. Extensive experiments across multiple public datasets demonstrate that our approach, by effectively incorporating person attribute group semantic information, achieves substantial performance improvements in ReID tasks, exhibiting superior generalization capabilities compared to existing frameworks.

Gait recognition is a biometric technology that distinguishes individuals by analyzing their walking patterns and holds significant potential for development. Current methods primarily focus on extracting gait features either from the overall appearance or specific local regions. However, they often overlook the partitioning of local regions based on the individual's body structure, as well as the weighted relationships between global and local features. We propose a Weighted Global-Local Feature Fusion Module to partition local features according to human body parts and adaptively integrate global and local gait features. This approach facilitates fine-grained learning of part-level local features and enhances the discriminative representation of gait features. Furthermore, we employ an Attention-based Multiscale Temporal Aggregation operation to adaptively fuse motion features from different time scales, preserving crucial spatio-temporal information while reducing the length of the time series. The average Rank-1 accuracy in CASIA-B and OUMVLP datasets is 93.0% and 90.7%, respectively. The experimental results demonstrate that our method achieves satisfactory recognition performance, indicating its potential for advancing gait recognition technology. (c) 2025 SPIE and IS&T

Due to inherent variability and unpredictability in real-world environments, single gait recognition models struggle to handle diverse covariate factors such as complex backgrounds, occlusions, viewpoint changes, and noise, often facing a trade-off between robustness and generalization. While ensemble learning offers a potential solution by integrating multiple models, traditional approaches often fail to manage heterogeneous knowledge transfer and mitigate negative transfer between conflicting experts. To address these challenges, we propose MEMLA-Gait, a novel framework for principled heterogeneous knowledge integration. Our key insight is that effective integration requires simultaneously resolving three interdependent issues: isolation between experts, varying reliability of expert knowledge, and superficial mimicry in passive transfer. MEMLA-Gait introduces three mechanisms: (1) a Mutual Learning Strategy (MLS) for collaborative knowledge refinement, (2) an Expert Authority Factor (EAF) that dynamically weights knowledge transfer based on discriminative power, and (3) an Adversarial Learning (AL) mechanism to promote deep feature assimilation. Extensive experiments on four challenging datasets show that MEMLA-Gait significantly outperforms state-of-the-art methods. Ablation studies further validate our framework, demonstrating that the unified system achieves substantially higher performance than any individual component, confirming the necessity of addressing all three challenges in concert.

Weakly supervised semantic segmentation (WSSS) has emerged as a promising approach to reduce annotation costs in autonomous driving applications. However, driving scenes present unique challenges that existing WSSS methods struggle to address, including complex spatial relationships, dramatic scale variations, and severe class imbalances. While recent CLIP-based methods have shown potential by leveraging vision-language alignments, they exhibit notable performance degradation when handling intricate urban environments, particularly struggling with small safety-critical objects and precise boundary delineation. We propose UASeg, a novel WSSS framework specifically designed to overcome these limitations through two key innovations: the Context Adaptive Decoder (CAD) and the Multi-Scale Uncertainty-Aware Adaptive Training (MUAT). CAD introduces a dynamic feature processing mechanism that adaptively modulates attention and convolution operations based on local scene complexity, enabling more effective representation of diverse object scales and spatial configurations. This is complemented by MUAT, which integrates uncertainty-guided learning with multi-scale analysis to enhance feature discrimination and address class imbalance issues. Through comprehensive experiments on the Cityscapes, CamVid, and WildDash2 datasets, we demonstrate that UASeg achieves state-of-the-art performance.

Low information entropy and coarse-grained motion pattern descriptions remain critical challenges for achieving accurate gait recognition in complex environments. To address these limitations, we propose FMCBGait, a novel fine-grained multimodal gait recognition framework that leverages the complementary strengths of gait parsing and skeleton modalities. In the parsing modality, we employ fine-grained segmentation of human semantic structures to extract body part patches of varying sizes, which are subsequently normalized to a unified scale using the Multi-Scale Adaptive Patch Embedding (MSAPE) module for effective feature encoding. For the skeleton modality, we construct a skeleton hypergraph to represent human topology, enabling the capture of higher-order interactions through a Hypergraph Convolutional Network (HGCN). To enhance inter-modal complementarity, representative visual cues from parsing data are embedded into the skeleton hypergraph, while structural cues from the skeleton modality are reciprocally embedded into the parsing modality. Additionally, the Multi-Stage Modal Fusion Strategy (MSMFS) integrates these modalities through early-stage cues embedding, mid-stage Body Part Distribution Guidance (BPDG), and late-stage Heterogeneous Non-Local (HNL) attention, facilitating comprehensive feature fusion. FMCBGait achieves state-of-the-art performance, with Rank-1 accuracies of 75.5% and 80.7% on two challenging datasets, demonstrating its robustness and efficacy for multimodal gait recognition.

Weakly Supervised Semantic Segmentation (WSSS) with image-level labels offers a promising solution to the expensive problem of pixel-level annotation. However, the prevalent use of Class Activation Maps (CAMs), while effective, often results in inaccurate object boundaries and poor edge details in generated pseudo- labels. To overcome these limitations, this paper presents a novel Multi-representation Fusion Learning (MFL) framework that leverages the remarkable capabilities of the Segment Anything Model (SAM) to enhance feature learning in WSSS. The MFL framework directly addresses the shortcomings of CAM-based pseudo-labels by incorporating rich semantic and edge information extracted from SAM. This is achieved through two dedicated modules: the Semantic-Guided Distilled Attention (SGDA) module and the Edge-Guided Distilled Attention (EGDA) module. These modules enable the network to learn more discriminative features by leveraging the SAM's knowledge, leading to higher-quality pseudo-labels. Furthermore, the proposed Multi-representation Fusion Module (MFM), based on a dual-layer routing attention mechanism, effectively fuses the semantic and edge features learned by the SGDA and EGDA, resulting in more refined pseudo-labels for training. Extensive experiments on PASCAL VOC 2012 and MS COCO 2014 datasets demonstrate that the MFL framework significantly outperforms existing WSSS methods, achieving state-of-the-art performance.

Use-After-Free vulnerabilities pose significant threats in memory-unsafe languages like C/C++, potentially leading to memory corruption, data leakage, or arbitrary code execution. This paper presents a novel UAF detection method based on Memory Tagging Extension, which enhances traditional virtual address-based defenses by introducing dynamic tag embedding and tag-state tracking. By binding a memory tag to both allocated objects and their referencing pointers, and modifying the tag upon deallocation, this approach enables efficient detection of dangling pointer dereferences through tag mismatches. The system integrates seamlessly with the Linux ptmalloc2 allocator using dynamic linking, allowing for high compatibility without requiring source code modifications. Experimental results on the Juliet Test Suite demonstrate a detection accuracy of 98%, while evaluations on the MiBench benchmark confirm significantly reduced runtime and memory overhead compared to existing tools such as ASan, QASan, and CETS. The proposed method balances security, performance, and engineering practicality, making it particularly suitable for resource-constrained and security-critical environments like embedded or IoT systems.

In multi-object tracking (MOT) tasks, maintaining long-term identity consistency of targets in complex scenes remains a challenging research problem. Traditional prediction methods based on visual appearance features and motion trajectories struggle to dynamically and continuously preserve the unique representation of targets in complex environments. This limitation leads to tracking drift and identity confusion when targets undergo occlusion, blurring, or changes in scene dynamics and motion patterns, significantly degrading tracking performance. To address this issue, we propose a novel approach, SemTG-Track, which links the same target through cross-modal semantic information. By integrating a vision-language model with a hybrid LoRA expert system, our method enhances tracking performance through fine-grained modality alignment and dynamic semantic matching strategies. The SemTG-Track framework consists of three core modules: Semantic-unit Temporal Completeness Generation (STCG), Heterogeneous Semantic Representation Alignment (HSRA), and Temporal Sampling and Dynamic Matching (TSDM). Specifically, the STCG module leverages a vision-language model to generate a rich semantic knowledge graph for targets, the HSRA module enhances the generalization capability of semantic units through a dual-domain expert semantic fusion mechanism, and the TSDM module improves the efficiency and accuracy of multi-object tracking via dynamic sampling and context-aware matching mechanisms. Experimental results demonstrate that the proposed method outperforms baseline approaches, achieving improvements of 2.0 and 4.1 percentage points in MOTA and HOTA, respectively, on the MOT17 dataset. On the MOT20 dataset, our method also achieves gains of 0.4 and 2.2 percentage points in MOTA and HOTA, respectively, validating the effectiveness of our approach.