NiVe1303
Description
NiVe1303 is a pure nighttime RGB vehicle re-identification (ReID) dataset collected from real roadside cameras. It contains 31,735 images of 1,303 vehicle identities, captured by an eight-camera RGB array mounted along a bidirectional three-lane urban road between 20:00 and 24:00. Four viewing directions (front, rear, left, and right) provide the cross-view appearance variation required for vehicle ReID. All samples depict moving vehicles on an open road, and therefore exhibit the imaging degradations characteristic of nighttime roadside surveillance: uneven illumination, headlight glare, motion blur, dark glass regions, and locally underexposed body contours. Each sample is annotated with vehicle identity, camera ID, and timestamp. In addition, NiVe1303 provides part-level masks for four semantic regions: front (front fascia plus a small area toward the front wheels), rear (defined analogously), top (hood, front and rear windshields, and roof), and side (doors, side windows, and wheels). 1,000 samples were manually annotated, and a mask annotator trained on these produced the remaining masks. The dataset is split into a training set of 17,070 images from 703 identities and a test set of 600 disjoint identities with 2,887 query and 11,778 gallery images at 256 x 256 resolution, following the evaluation protocol of DukeMTMC4ReID. License plates and faces are not visually discernible due to motion blur. Data collection and release were approved by the local authority. NiVe1303 supports research on nighttime vehicle ReID, part-based representation learning, and low-light image enhancement for recognition tasks.
Files
Institutions
- Beijing University of TechnologyBeijing, Beijing