EVENT DOSSIER
arXiv:2608.31107v2: VeriCam is released to address the shortcomings of basic models in fine-grained classification of unknown categories. This study utilizes image verification tasks to construct a feature space with detailed features and establishes relationships between data points by training the model to distinguish pairs of images of the same and different classes. VeriCam employs a naive algorithm and a specific configuration of the Leiden graph clustering algorithm for graph clustering, and is validated on the LPLCv2 traffic monitoring dataset. To overcome the inherent challenges of camera bias in this dataset, VeriCam dynamically identifies cameras in a label-independent manner, constructing a fair and unbiased benchmark. In cross-device scenarios, its validation baseline F1-Score reaches 93.45, and the clustering V-Measure score is 80.13. All code has been made available in the GitHub repository.
arXiv:2608.31107v2 提出 VeriCam，旨在解决基础模型在细粒度未知类别分类上的不足。该流程利用图像验证任务构建包含精细细节的特征空间，并通过训练模型区分同异类图像对来建立数据点间的关系图。研究采用朴素算法及 Leiden 图聚类算法的特定配置进行图聚类，并在 LPLCv2 交通监控数据集上进行验证。针对该数据集固有的拍摄设备偏差挑战，VeriCam 以标签无关方式动态识别拍摄设备，构建了公平无偏基准。在跨设备场景中，验证基线 F1-Score 达 93.45，聚类步骤 V-Measure 得分为 80.13。所有代码已公开至 https://github.com/lmlwojcik/VeriCam。