Numerous studies leverage pose estimation to extract human keypoint data and then classify sitting postures. However, employing neural networks for direct keypoint classification often yields suboptimal results. Alternatively, modeling keypoints into other data representations before classification introduces redundant information and substantially increases inference time. In addition, most existing methods perform well only under a single fixed viewpoint, limiting their applicability in complex real-world scenarios involving unseen viewpoints. To better address the first limitation, we propose KeypointNet, which employs a decoupled feature extraction strategy consisting of a Keypoint Feature Extraction module and a Multi-Scale Feature Extraction module. In addition, to enhance multi-view recognition capability, we propose the Multi-View Simulation (MVS) algorithm, which augments the viewpoint information by first rotating keypoints and then repositioning the camera. Simultaneously, we propose the multi-view sitting posture (MVSP) dataset, designed to simulate diverse real-world viewpoints. The experimental results demonstrate that KeypointNet outperforms the other state-of-the-art methods on both the proposed MVSP dataset and the other public datasets, while maintaining a lightweight and efficient design. Ablation studies demonstrate the effectiveness of MVS and all KeypointNet modules. Furthermore, additional experiments highlight the superior generalization, small-sample learning capability, and robustness to unseen viewpoints of KeypointNet.

Temporal action localization (TAL) aims to predict action instance categories in videos and identify their start and end times. However, existing Transformer-based backbones focus only on global or local features, resulting in the loss of information. In addition, both global and local self-attention mechanisms tend to average embeddings, thereby reducing the preservation of critical features. To solve these two problems better, we propose two kinds of attention mechanisms, namely multi-headed local self-attention (MLSA) and max-average pooling attention (MA) to extract simultaneously local and global features. In MA, max-pooling is used to select the most critical information from local clip embeddings instead of averaging embeddings, and average-pooling is used to aggregate global features. We use MLSA for modeling local temporal context. In addition, to enhance collaboration between MA and MLSA, we propose the double attention block (DABlock), comprising MA and MLSA. Finally, we propose the final network double attention network (DANet), composed of DABlocks and other advanced blocks. To evaluate DANet’s performance, we conduct extensive experiments for the TAL task. Experimental results demonstrate that DANet outperforms the other state-of-the-art models on all datasets. Finally, ablation studies demonstrate the effectiveness of the proposed MLSA and MA. Compared with structures using backbone with convolution and global Transformer, DABlock consisting of MLSA and MA has a superior performance, achieving an 8% and 0.5% improvement on overall average mAP, respectively.

Pseudo-label-based deep clustering has emerged as a compelling paradigm for unsupervised visual learning, yet two fundamental limitations impede its effectiveness: (1) pseudo-label noise, where unreliable hard assignments generated in early training accumulate and propagate errors across iterations; and (2) confidence-insensitive pseudo-label supervision, where one-hot pseudo-labels are treated with equal importance despite substantial differences in their reliability, increasing the influence of uncertain samples, particularly those near cluster boundaries. To address both issues simultaneously, we propose PSPLC—a Progressive Sample-Weighted Pseudo-Label Clustering framework with three synergistic modules. First, a MoCo-based contrastive pre-training module extracts discriminative feature representations. Second, a progressive sample-weighted pseudo-label training module (the core contribution) iterates an expectation step—which progressively expands a high-confidence sample set outward from calibrated cluster centres and assigns one-hot pseudo-labels with continuous confidence weights derived from a global–local confidence mechanism—with a maximisation step that jointly optimises a clustering consistency loss, a confidence-weighted one-hot pseudo-label loss, and an entropy balance loss. Third, a semi-supervised boosting module (PSPLC-Boost) identifies reliable samples via nearest-neighbour semantic consistency and jointly fine-tunes the feature extractor and clustering head with a FixMatch-inspired strategy. Extensive experiments on five benchmarks (STL-10, CIFAR-10, CIFAR-20, ImageNet-10, ImageNet-Dogs-15) show that PSPLC achieves state-of-the-art performance among purely unsupervised methods, and PSPLC-Boost surpasses all competitors on all five datasets across all three metrics.

Diffusion models have shown strong promise for neural combinatorial optimization, yet existing approaches define the forward noising process over a fully connected graph, discarding the sparse topological structure inherent to the Traveling Salesman Problem (TSP), and their multi-step inference incurs prohibitive latency. We address these two limitations through a pair of complementary models. First, we propose GCDTSP, a graph-constrained diffusion model that replaces the topology-agnostic uniform transition matrix with an instance-dependent graph Laplacian-based heat-kernel transition matrix, so that the evolution of the edge-valued state is governed by the topology of each TSP instance; a feasibility constraint loss combining degree conservation, symmetry, and locality priors further guides the GatedGCN denoising network toward valid Hamiltonian cycles. GCDTSP outperforms autoregressive baselines and prior diffusion solvers under greedy decoding, and generalizes zero-shot to real-world TSPLIB instances. Building upon GCDTSP, we introduce GCDTSP-D, a progressive distillation framework that iteratively halves the required denoising steps via teacher–student parameter inheritance, combined with a cosine noise schedule that concentrates inference capacity in the constraint-sensitive low-noise regime. GCDTSP-D attains an approximately 37× speedup on TSP50 with a favorable quality–speed trade-off, and on the larger TSP500 benchmark even surpasses its teacher in both solution quality and speed.

Recent studies have proposed post-processing strategies that iteratively reconstruct a partial segment of current solutions using Neural Combinatorial Optimization (NCO) solvers, thereby improving their performance on large-scale Vehicle Routing Problem (VRP) instances. However, the reconstruction subproblem is essentially a Shortest Hamiltonian Path Problem (SHPP) instance, which differs fundamentally from the original VRP variant on which the solver is trained. Consequently, the NCO solver may suffer from performance degradation during reconstruction due to limited generalization across problem variants. Moreover, relying solely on the solver may be insufficient to effectively identify and correct complex intersections or suboptimal topological structures. To address these limitations, we propose a post-processing strategy termed Hybrid Reconstruction Framework (HRC). Specifically, HRC first fine-tunes the NCO solver on SHPP instances and then exploits the enhanced solver to perform large-neighborhood random reconstruction. Subsequently, HRC conducts small-neighborhood reconstruction using 2-opt and kNN-DGR. The experimental results on both synthetic and real-world Traveling Salesman Problem and Capacitated Vehicle Routing Problem instances demonstrate that HRC substantially improves the performance of two representative NCO solvers on large-scale instances and achieves better overall performance than the state-of-the-art reconstruction strategy. Finally, ablation studies further validate the effectiveness of all proposed designs.