ICPR 2026 Competition - TVRID

TVRID (Top-View RGB-Depth Person Re-Identification Dataset) is a benchmark for privacy-preserving person Re-ID in overhead camera settings, featuring 88 identities recorded by four synchronized Intel RealSense D455 cameras providing paired RGB and Depth streams. Each subject is observed twice per camera (IN/OUT) along a short path designed to induce systematic viewpoint and elevation changes (flat ground, ascent, descent, and an oblique roof view), creating realistic cross-camera distribution shifts within short 2–3 s observation windows. The dataset is distributed in two complementary formats: Original (the base dataset with the full raw frames, 640×480 RGB/Depth), and Extracted (a derived version where pedestrian-centered crops are automatically extracted from the Original frames to provide ready-to-use inputs around the person). GitHub : https://github.com/RaphaelDel/ICPR-TVRID