SortCS is a 'Multiple Object Tracker' as described in this paper, implemented in C#.
SORT is a barebones implementation of a visual multiple object tracking framework based on rudimentary data association and state estimation techniques. It is designed for online tracking applications where only past and current frames are available and the method produces object identities on the fly. While this minimalistic tracker doesn't handle occlusion or re-entering objects its purpose is to serve as a baseline and testbed for the development of future trackers.
SORT was initially described in this paper. At the time of the initial publication, SORT was ranked the best open source multiple object tracker on the MOT benchmark.
using SortCS;
ITracker tracker = new SortTracker(iouThreshold = 0.3f, maxMisses = 3);
tracker.Track(new[]
{
new RectangleF(1695,383,159,343),
new RectangleF(1293,455,83,213)
});
tracker.Track(new[]
{
new RectangleF(1699,383,159,341),
new RectangleF(1293,455,83,213)
});
tracker.Track(new[]
{
new RectangleF(1697,383,159,343),
new RectangleF(1293,455,83,213)
});
var tracks = tracker.Track(new[]
{
new RectangleF(1695,383,159,343),
new RectangleF(1293,455,83,213)
});
Assert.AreEqual(2 tracks.Count());
Assert.AreEqual(TrackState.Active, tracks.First().State);
Assert.AreEqual(4, tracks.First().History.Count);The performance of this implementation can be evaluation using the 'SortCS.Evaluate' Console Application. The output can be used for https://motchallenge.net/ and their TrackEval SDK. Brief instructions:
- Clone this repo and the TrackEval in the same folder / next to each other
- Run the SortCS.Evaluate Console app. The
--data-folderarguments needs to point to the data folder in theTrackEvalrepo. - Make sure that the outputs (tracks+detections) will be stored in
TrackEval/data/trackers/mot_challenge/MOT20-train/SortCS/data - Run TrackEval (according to their readme):
python scripts/run_mot_challenge.py --BENCHMARK MOT20 --SPLIT_TO_EVAL train --TRACKERS_TO_EVAL SortCS --METRICS HOTA CLEAR Identity VACE --USE_PARALLEL False --NUM_PARALLEL_CORES 1
Preliminary results for SortCS & MOT20 Train:
All sequences for SortCS finished in 103.38 seconds
HOTA: SortCS-pedestrian HOTA DetA AssA DetRe DetPr AssRe AssPr LocA RHOTA HOTA(0) LocA(0) HOTALocA(0)
MOT20-01 97.52 97.299 97.75 98.397 98.397 98.616 98.616 96.449 98.067 100 96.333 96.333
MOT20-02 76.905 97.8 60.478 98.736 98.737 65.202 86.718 96.909 77.271 78.464 96.803 75.955
MOT20-03 66.813 96.08 46.481 97.64 97.641 48.266 84.487 96.139 67.367 68.979 95.845 66.113
MOT20-05 64.614 96.537 43.255 98.026 98.03 46.88 79.206 96.548 65.114 66.432 96.217 63.919
COMBINED 67.705 96.586 47.471 98.023 98.025 50.691 82.04 96.483 68.213 69.6 96.196 66.953
CLEAR: SortCS-pedestrian MOTA MOTP MODA CLR_Re CLR_Pr MTR PTR MLR sMOTA CLR_TP CLR_FN CLR_FP IDSW MT PT ML Frag
MOT20-01 100 96.333 100 100 100 100 0 0 96.333 19870 0 0 0 74 0 0 0
MOT20-02 99.806 96.819 99.941 99.97 99.971 100 0 0 96.626 154695 47 45 208 270 0 0 32
MOT20-03 98.894 95.925 99.616 99.807 99.808 100 0 0 94.827 313054 604 601 2264 702 0 0 533
MOT20-05 98.905 96.355 99.557 99.777 99.78 99.914 0.085543 0 95.268 644900 1444 1419 4216 1168 1 0 1102
COMBINED 99.044 96.299 99.633 99.815 99.818 99.955 0.045147 0 95.35 1132519 2095 2065 6688 2214 1 0 1667
Identity: SortCS-pedestrian IDF1 IDR IDP IDTP IDFN IDFP
MOT20-01 100 100 100 19870 0 0
MOT20-02 71.141 71.14 71.141 110084 44658 44656
MOT20-03 60.351 60.351 60.352 189296 124362 124359
MOT20-05 57.143 57.142 57.144 369332 277012 276987
COMBINED 60.689 60.689 60.69 688582 446032 446002
VACE: SortCS-pedestrian SFDA ATA
MOT20-01 96.356 100
MOT20-02 96.795 58.253
MOT20-03 95.943 31.567
MOT20-05 96.295 32.092
COMBINED 96.359 35.234
Count: SortCS-pedestrian Dets GT_Dets IDs GT_IDs
MOT20-01 19870 19870 74 74
MOT20-02 154740 154742 403 270
MOT20-03 313655 313658 2096 702
MOT20-05 646319 646344 3547 1169
COMBINED 1134584 1134614 6120 2215
- MaartenX, for implementing some of the core pieces of this repo
- abewley's Python implementation of Sort, used for validation/reference