Closed Reeshark closed 5 years ago
Hi, thanks for trying out this code. It is because MOT16 groundtruth includes some extra non-people classes like walls, etc. Thus there would be 'FP' in evaluation results. For example, use the gt.txt as detection results of MOT16-02-DPM. Then there would be 3402 fps. Which is expected and reasonable.
[GT PREPROCESSING]: Removing non-people classes, remaining 26601/30003 boxes
[TRACK PREPROCESSING]: remove distractors and low visibility boxes, remaining 21983/30003 computed boxes
('Distractors:', array([ 1., 5., 6., 11., 12., 13., 24., 25., 30., 45., 46.,
47., 65., 66.]))
[GT PREPROCESSING]: Removing distractor boxes, remaining 18581/26601 computed boxes
('\n', '******************************', 'MOT17-02 Evaluation', '******************************')
IDF1 IDP IDR | Rcll Prcn FAR | GT MT PT ML | FP FN IDs FM | MOTA MOTP MOTAL
91.6 84.5 100.0 | 100.0 84.5 5.67 | 62 62 0 0 | 3402 0 0 0 | 81.7 100.0 81.7
('\n', '******************************', 'Summary Evaluation', '******************************')
IDF1 IDP IDR | Rcll Prcn FAR | GT MT PT ML | FP FN IDs FM | MOTA MOTP MOTAL
91.6 84.5 100.0 | 100.0 84.5 5.67 | 62 62 0 0 | 3402 0 0 0 | 81.7 100.0 81.7
Thanks for reply, I have realized this issue. Your explaination is same as my conclusion.
I have read your paper "Tracklet Association Tracker: An End-to-End Learning-based Association Approach for Multi-Object Tracking"
I have read your paper "Tracklet Association Tracker: An End-to-End Learning-based Association Approach for Multi-Object Tracking"
Thank you for your attention. I hope the paper and this code can be helpful for you. FYI, I updated the code to python3 today and polish the coding styles by linter.
Why the MOTA and IDF1 are not 100% when I input the ground truth as the tracking results? For example your data MOT16-13, which MOTA is only 31.8 and IDF1 is only 74.6?