MARS Person Re-ID Video Dataset

MARS stands for Motion Analysis and Re-identification Set. It is a large-scale pedestrian re-identification video dataset and an extension of the Market-1501 dataset.

The data is collected by six nearly synchronized cameras. The dataset contains 1,261 different pedestrians, each captured by at least 2 cameras. The differences in pedestrians' poses, clothing colors, and lighting, as well as poor image quality, make it difficult to produce high matching accuracy. In addition, the dataset contains 3,248 distractors to make it more realistic. The DPM model and GMMCP tracker are used to automatically generate short trajectories (mostly 25-50 frames long).