When the Transformer based object tracking method employs the global attention mechanism to model the spatial relations between the search area and the template, the target deformation can lead to the degradation of feature discriminability, causing confusion between the target and the background. To solve this problem, a dynamic semantic clustering relation modeling method for object tracking is proposed. First, a semantic relation modeling module is constructed. Local attention mechanisms in the feature space are employed to concentrate on semantically similar feature vectors, thereby effectively suppressing erroneous interactions between the target and the distracting background. Second, graph neural networks are utilized to capture local correlations and design a dynamic semantic clustering module. The module adaptively generates semantic category partitions, enabling dynamic attention mechanisms to enhance the discriminative information between the target and the background. Finally, a semantic background elimination strategy is designed to effectively suppress the interference from background features during relationship modeling, thereby improving tracking efficiency. Experimental results on six benchmark datasets demonstrate the superiority of the proposed method.

In modern data analysis and machine learning applications, how to extract critical information from newly acquired data for efficient grouping (clustering) and annotation remains a central challenge for clustering algorithms. Traditional unsupervised clustering algorithms struggle to meet the high-quality data requirements of complex tasks, such as pre-training large models due to the lack of prior information guidance. Active learning methods can effectively improve clustering accuracy, but their practical application is constrained by high human interaction costs and computational overhead. To address these issues, an algorithm of active clustering with tailored nearest neighbor graph(ACNNG) is proposed. A sparse nearest neighbor graph is constructed to model relationships between data points. Based on this graph, the topological centrality and uncertainty of data points are comprehensively computed to effectively identify key data points. A small number of pairwise constraints are collected from users to significantly enhance clustering accuracy. Furthermore, an efficient label propagation mechanism collaborating with the nearest neighbor graph structure is employed. By leveraging the sparse graph structure for low-cost label propagation, ACNNG substantially reduces the spatiotemporal complexity and improves scalability for large-scale data processing. Experiments on real-world and synthetic datasets demonstrate that ACNNG achieves higher clustering accuracy with fewer pairwise constraints, less runtime, and less memory consumption, showcasing its potential for practical applications.