Omni Scale Network for Video-Based Person Re-Identification Against Unconstrained Conditions

Person re-identification (Re-ID) is a computer vision field incorporating deep learning techniques. The objective is to efficiently and accurately associate, detect, and track an individual's identification across various camera views or locations within a video or image sequence. An advancement over image-based, video-based re-identification is far more difficult and intricate. Distractions in both space and time known as Spatial and Temporal distractors such as occlusions, lighting, background clutter, and position changes, contribute to the complexity of this task. With the growing need for intelligent video surveillance, video-based re-id plays a crucial role in real-world surveillance systems for security and tracking. This paper proposes a CNN based omni scale network model for video-based person re-identification that reidentifies the person in unconstrained conditions and deals with discriminative and generalizable features. Extensive studies show the efficacy of the suggested approach, and the aggregate results on the PRID2011 and iLIDS-VID dataset give outstanding performance by a significant margin and achieved 90.8% (+0.2) rank-1 accuracy, 98.6%(+0.3) rank-5 accuracy, 100% rank-10 accuracy, 100% rank-20 accuracy and mean average precision of 90.7% on iLIDS-VID dataset and 96.8%(+0.2) rank-1 accuracy, 98.9% rank-5 accuracy, 100% rank-10 accuracy, 100% rank-20 accuracy and mean average precision of 92% on PRID2011 dataset.