ReID-X: Scalable Re-ID for Dynamic Camera Systems 2026 is a multi-camera detection and person re-identification (ReID) challenge designed to build robust systems for dynamic environments.
Participants are given video sequences captured from multiple synchronized cameras. The core task is to assign consistent global identities (pred_track_id) to each individual across all frames and all cameras within a sequence. The goal extends beyond accurate localization; it requires maintaining identity consistency over time and across different viewpoints. This involves handling complex real-world scenarios where individuals may appear in different viewpoints, disappear due to occlusion, and reappear later under varying lighting and pose conditions.
Top-performing participants may be considered for internships and collaboration opportunities on INICAI-driven projects.
Submissions are evaluated using HOTA (Higher Order Tracking Accuracy), a metric specifically designed for multi-object tracking that jointly measures detection quality and identity association quality. HOTA is averaged over IoU thresholds from 0.5 to 0.95. The score ranges from 0.0 to 1.0, with a naive baseline typically scoring between 0.15 and 0.25.
The metric rewards: * Accurate bounding box localization (DetA) * Consistent identity assignment across frames and cameras (AssA) * Stable IDs over time without unnecessary switches * Complete trajectories without fragmentation * Correct separation of different individuals
Rankings are determined by the Private Leaderboard, which scores the remaining ~70% of test data, while the Public Leaderboard scores approximately 30%.
For each detection in sample_submission.csv, participants must assign a pred_track_id and predicted bounding box. The file must contain exactly the following columns: row_id, frame_id, pred_track_id, xmin, ymin, xmax, ymax. Key requirements include ensuring pred_track_id is an integer, every row_id appears exactly once, and bounding boxes are valid. Crucially, the same person must have the same pred_track_id across all frames and cameras.