Open voyagerw opened 4 years ago
I also tested real videos with 'mot17_fulltrain' model, the ID switches immediately as long as a person is covered by other persons or objects on the two neighboring frames, or to say the detection of this person is interrupted.
@ctrcv the matching strategy in this paper is to throw away a tracked object if it is missed in one frame(i.e. no cross-frame matching) which works fine for the few datasets (MOT, KITTI, nuScene). You may want to consider born/death strategy(e.g. 3 continuous hits initialize a tracklet and only delete it after 2 or 3 misses) here if you want to work on other datasets.
I tried to test in test sequences with mot17_fulltrain, but the results were not so good enough. As long as a person was covered by other persons on the two neighboring frames , his ID would be changed immediately.
The result images of tracking were generated by command '--debug 4'. I was wondering if there was something wrong with my operation