v6.0
版本发布时间: 2022-06-10 01:36:43
mikel-brostrom/boxmot最新发布版本:v11.0.3(2024-10-01 16:26:21)
Important updates
StrongSORT implemented (https://arxiv.org/pdf/2202.13514.pdf)
- stronger appearance descriptor (OSNet)
- camera motion compensation (ECC)
- NSA Kalman filter (NSA)
- EMA feature updating mechanism (EMA)
- matching with motion cost (MC)
- abandone matching cascade in favor of a vanilla global linear matching (woC)
Distance metric changed to cosine according to https://github.com/KaiyangZhou/deep-person-reid/issues/502
Complete track.py arguments refactor
MOT16 Train evaluation results
Relevant changed/used hparams: imgz 1280, crowdhuman_yolov5m, OSNet_x0_25_msmt17, StrongSORT. Notice that none of the models used during the evaluation has ever seen any of the MOT16 data and that our object detection model is a modest Yolov5m.
HOTA: kf-pedestrian HOTA DetA AssA DetRe DetPr AssRe AssPr LocA RHOTA HOTA(0) LocA(0) HOTALocA(0)
MOT16-02 36.417 36.273 36.741 38.196 76.251 39.184 77.59 81.577 37.433 46.154 77.361 35.705
MOT16-04 59.324 58.508 60.726 63.282 76.509 64.924 79.5 81.075 61.927 78.835 75.37 59.418
MOT16-05 40.098 37.281 43.23 39.956 74.391 51.86 69.098 81.876 41.554 51.282 77.049 39.512
MOT16-09 53.181 57.122 49.57 65.872 74.664 54.868 78.804 85.721 57.137 65.772 82.193 54.06
MOT16-10 49.705 49.813 49.762 53.696 74.831 53.942 77.666 80.652 51.686 65.955 75.773 49.976
MOT16-11 62.896 59.81 66.362 71.318 73.696 72.965 82.899 86.879 68.784 74.954 83.531 62.61
MOT16-13 45.203 41.504 49.567 44.959 72.74 54.969 74.378 80.493 47.163 59.49 75.279 44.784
COMBINED 53.005 50.992 55.641 55.611 75.463 60.424 79.256 81.937 55.531 68.581 76.857 52.709
CLEAR: kf-pedestrian MOTA MOTP MODA CLR_Re CLR_Pr MTR PTR MLR sMOTA CLR_TP CLR_FN CLR_FP IDSW MT PT ML Frag
MOT16-02 43.56 78.641 44.451 47.272 94.369 16.667 53.704 29.63 33.463 8430 9403 503 159 9 29 16 377
MOT16-04 69.786 78.669 69.956 76.334 92.289 49.398 34.94 15.663 53.503 36302 11255 3033 81 41 29 13 582
MOT16-05 43.018 79.285 43.825 48.768 90.797 17.6 62.4 20 32.916 3325 3493 337 55 22 78 25 210
MOT16-09 61.29 84.193 62.165 75.195 85.231 52 48 0 49.404 3953 1304 685 46 13 12 0 108
MOT16-10 61.154 77.435 61.788 66.772 93.054 29.63 55.556 14.815 46.087 8225 4093 614 78 16 30 8 558
MOT16-11 63.56 85.508 63.942 80.358 83.037 53.623 36.232 10.145 51.915 7372 1802 1506 35 37 25 7 145
MOT16-13 50.629 77.098 51.397 56.603 91.578 26.168 44.86 28.972 37.666 6481 4969 596 88 28 48 31 291
COMBINED 60.025 79.394 60.516 67.104 91.06 32.108 48.549 19.342 46.198 74088 36319 7274 542 166 251 100 2271
Identity: kf-pedestrian IDF1 IDR IDP IDTP IDFN IDFP
MOT16-02 47.717 35.81 71.488 6386 11447 2547
MOT16-04 75.047 68.56 82.891 32605 14952 6730
MOT16-05 54.752 42.08 78.345 2869 3949 793
MOT16-09 65.124 61.29 69.47 3222 2035 1416
MOT16-10 66.758 57.331 79.896 7062 5256 1777
MOT16-11 74.053 72.858 75.287 6684 2490 2194
MOT16-13 60.474 48.926 79.158 5602 5848 1475
COMBINED 67.195 58.357 79.189 64430 45977 16932
Count: kf-pedestrian Dets GT_Dets IDs GT_IDs
MOT16-02 8933 17833 146 54
MOT16-04 39335 47557 139 83
MOT16-05 3662 6818 136 125
MOT16-09 4638 5257 57 25
MOT16-10 8839 12318 122 54
MOT16-11 8878 9174 164 69
MOT16-13 7077 11450 135 107
COMBINED 81362 110407 899 517
Performance boosts from updating DeepSORT to StrongSORT
Bug fixes
Confidences are now extracted correctly: https://github.com/mikel-brostrom/Yolov5_DeepSort_OSNet/issues/375