Current Vision Language Models (VLMs) have demonstrated remarkable proficiency in general visual understanding. However, they encounter a critical perception bottleneck when applied to Training-free Visual Search, particularly for fine-grained or small objects. Relying on static, low-resolution holistic encoding, standard VLMs often fail to distinguish actual visual evidence from hallucinations driven by language priors. To address this, we introduce GazeSim, a novel training-free framework that transforms VLMs into active visual agents capable of mimicking human search behaviors. Drawing inspiration from cognitive processes, GazeSim operates on a "Look-Plan-Verify" paradigm. First, we propose a Visual-Aligned Context Retrieval mechanism that selects layout-aware demonstrations to implicitly teach the model scene-specific search strategies. Second, the VLM acts as an agentic planner to generate dynamic scanpaths governed by biological constraints, such as saccadic amplitude and inhibition of return. Finally, we implement Foveal Verification, where the model scrutinizes high-resolution local crops along the simulated trajectory to perform coarse-to-fine joint reasoning. Extensive experiments on the COCO-Search18 and COCO-FreeView demonstrate that GazeSim establishes a new state-of-the-art in visual search, achieving a significant accuracy improvement over existing baselines.

Fine-Grained Visual Classification (FGVC) is essential for enhancing the accuracy and robustness of computer vision systems across diverse real-world applications. However, existing FGVC methods struggle in open-set scenarios, where models must not only classify known categories with high precision but also effectively detect and reject unknown classes. To address these challenges, we propose the Multi-Stream Collective Deliberation Network (MSCDNet), which introduces a novel Collective Deliberation Attention (CoDA) module to collaboratively extract discriminative features while mitigating attention collapse, ensuring a more comprehensive focus on fine-grained details. Furthermore, a contrastive learning mechanism is integrated to enhance feature separability and improve the system’s robustness against unknown categories. Extensive experiments on two benchmark datasets, CUB-200-2011 and FGVC-Aircraft, demonstrate that MSCDNet consistently outperforms state-of-the-art methods in both closed-set and open-set recognition tasks. Visualization results further highlight MSCDNet’s ability to focus on critical and distinguishing features, leading to significant improvements in classification performance.

The issue of how to identify the liability of subjects after a traffic accident takes place remains a puzzle regarding the SAE classification system. The SAE system is not good at dealing with the problem of responsibility evaluation; therefore, building a new classification system for self-driving cars from the perspective of the subject's liability is a possible way to solve this problem. This new system divides automated driving into three levels: i) assisted driving based on the will of drivers, ii) automated driving based on the will of the manufacturers, and iii) fully automated driving based on social will. The corresponding responsible subjects for level one, level two, and level three are the drivers, the manufacturers, and society, respectively.