MARS-Gait

MARS-Gait is a new outdoor gait recognition dataset. It is built based on the MARS (Motion Analysis and Re-identification Set) dataset, whose samples are collected in the campus of Tsinghua university. MARS is used for video-based person Re-ID task, which uses 6 near- synchronized cameras located in the campus for recording the pedestrians. It provides the tracklets with person and camera IDs as the annotations. To obtain the gait recognition dataset, we first screened all samples, removing the tracklets where people are obviously not walking, such as cycling or sitting. After that, we adopt an instance segmentation algorithm HTC, to extract silhouettes from original RGB sequences and get the corresponding gait sequences. The person ID of each sample is inherited from that of the original tracklet. In total, MARS-Gait consists of 1,246 subjects and 16,093 gait sequences, of which 616 subjects with 8,101 sequences are used for training and the rest 630 subjects with 7,992 sequences for test.