A full-body human crop and triplet dataset for person re-identification derived from MOT20 crowded-scene sequences

This repository contains a derived person re-identification dataset generated from four MOT20 training sequences: MOT20-01, MOT20-02, MOT20-03, and MOT20-05. The release includes filtered full-body pedestrian crop images, metadata files, identity mapping files, export statistics, and triplet annotations for metric-learning-based person re-identification research.

A total of 1,336,920 original MOT20 ground-truth pedestrian annotations were examined. After removing zero-confidence annotations and applying a visibility threshold of 0.5, 617,655 candidate instances remained. Subsequent crop-size filtering and pose-based full-body validation reduced this set to 387,426 final exported crop images. The final release contains 310,935 training images, 36,603 validation images, 465 query images, and 39,423 gallery images, covering 2,089 exported identities.

The repository also contains triplet annotations in JSONL format. In total, 310,602 triplets are included: 277,688 training triplets, 32,466 validation triplets, and 448 test triplets. Approximately 80% of triplets are marked as hard negatives. The dataset is intended for person re-identification, metric learning, retrieval evaluation, and hard-negative mining studies in crowded-scene conditions.

Exported crop images follow the naming convention _c_f.jpg, where personID is the identity label, cameraID is the ReID-style camera tag used in the export, and frameID is the source frame number.

This dataset is derived from the public MOT20 benchmark and is distributed as a processed research resource for reuse in ReID pipelines.