I-PRIDe: Video-based Person Re-Identification

The surveillance of infrastructures, strategic assets, and public places is a critical capability in preventing crimi-nal activities. The possibility of re-identifying recurrent actors contributes to severity assessment and activation of appropriate countermeasures. This paper presents a novel dataset and a baseline analysis for video-based person re-identification (Re-ID). Our baseline addresses three issues: the performance of OsNet [1] as a feature extractor, temporal modeling mechanisms to improve the overall Re-Idaccuracy, and an assessment of the effectiveness of the selected methods through different strategies in selecting frames from the videos for the temporal models. We make three distinct contributions. First, we introduce a new dataset I-PRIDe 1 1 The dataset is available at https://datasets.vicomtech.org/di24-i-pride/ readme.txt that mitigates the lack of acquisition angles and poses variety in existing video-based person Re-Iddatasets. It counts with indoor and outdoor scenes, multiple camera angles (0°, 30°, 60°), and variety in the pose of the involved subjects. It contains 104 identities within 2,452 duly annotated sequences. Second, we justify that OsNet is well generalizable yet sufficiently discriminative to adapt to various perspectives and poses without fine-tuning. Third, we derive insights into evaluating the performance of temporal modeling techniques with specific frame selection settings. Notably, the rank-1 accuracy of our method reaches 89.86%, and the mean average precision achieves 85.71 %.