| Configuration | FPS (osnet_x1 GTX1650) | FPS (mobilenetv2_x1_0 GTX1650) | FPS (osnet_x0_25 CPU) | FPS (osnet_x1 T4) | FPS (ResNet50 TX2) |
|---|---|---|---|---|---|
| StrongSORT | 4.85 | 5.09 | 1.33 | 5.80 | 0.96 |
| Fast-StrongSORT (iou_threshold=0.0) | 5.23 (+8%) | 5.33 (+5%) | 1.68 (+26%) | 6.67 (+15%) | 1.17 (+22%) |
| Fast-StrongSORT (iou_threshold=0.1) | 5.59 (+15%) | 5.57 (+9%) | 2.02 (+52%) | 7.22 (+24%) | 1.40 (+46%) |
| Fast-StrongSORT (iou_threshold=0.2) | 5.97 (+23%) | 5.86 (+15%) | 2.48 (+86%) | 7.71 (+33%) | 1.73 (+80%) |
| Fast-StrongSORT (iou_threshold=0.3) | 6.42 (+32%) | 6.11 (+20%) | 2.94 (+121%) | 8.32 (+43%) | 2.07 (+116%) |
| Fast-StrongSORT (iou_threshold=0.4) | 6.84 (+41%) | 6.34 (+25%) | 3.64 (+174%) | 8.80 (+52%) | 2.52 (+163%) |
| Fast-StrongSORT (iou_threshold=0.5) | 7.23 (+49%) | 6.57 (+29%) | 4.51 (+239%) | 9.22 (+59%) | 3.21 (+234%) |
| Tracker | PDE* ↓ | FET** (s) ↓ | FPS ↑ | HOTA ↑ | AssA ↑ | IDF1 ↑ |
|---|---|---|---|---|---|---|
| StrongSORT | 100.0 | 2457 | 0.96 | 69.54 | 73.29 | 82.17 |
| FSS 0.0 | 76.38 (-24%) | 1946 (-21%) | 1.17 (+22%) | 69.36 (-0.26%) | 72.92 (-0.50%) | 81.88 (-0.35%) |
| FSS 0.1 | 58.67 (-41%) | 1598 (-35%) | 1.40 (+46%) | 69.39 (-0.22%) | 72.93 (-0.49%) | 81.81 (-0.44%) |
| FSS 0.2 | 43.93 (-56%) | 1241 (-50%) | 1.73 (+80%) | 69.61 (+0.1%) | 73.44 (+0.21%) | 82.56 (+0.47%) |
| FSS 0.3 | 33.40 (-67%) | 995 (-60%) | 2.07 (+116%) | 69.09 (-0.65%) | 72.50 (-1.07%) | 81.97 (-0.24%) |
| FSS 0.4 | 23.46 (-77%) | 772 (-69%) | 2.52 (+163%) | 68.31 (-1.77%) | 70.87 (-3.30%) | 80.91 (-1.53%) |
| FSS 0.5 | 16.14 (-84%) | 547 (-78%) | 3.21 (+234%) | 68.86 (-0.98%) | 71.88 (-1.92%) | 81.21 (-1.17%) |
*PDE is the abbreviation of Percentage of Detections with Extraction. **FET is the abbreviation of Total Feature Extraction Time.
| Tracker | PDE* ↓ | FET** (s) ↓ | FPS ↑ | HOTA ↑ | AssA ↑ | IDF1 ↑ |
|---|---|---|---|---|---|---|
| StrongSORT | 100.0 | 14054 | 1.56 | 56.61 | 41.12 | 55.91 |
| FSS 0.0 | 85.34 (-15%) | 11972 (-15%) | 1.79 (+15%) | 57.07 (+0.81%) | 41.78 (+1.61%) | 56.35 (+0.77%) |
| FSS 0.1 | 73.40 (-27%) | 10518 (-25%) | 2.00 (+28%) | 56.55 (-0.11%) | 41.01 (-0.27%) | 55.90 (-0.03%) |
| FSS 0.2 | 58.64 (-41%) | 8548 (-39%) | 2.38 (+53%) | 56.80 (+0.34%) | 41.36 (+0.60%) | 56.19 (+0.49%) |
| FSS 0.3 | 43.02 (-57%) | 6434 (-54%) | 2.97 (+90%) | 55.46 (-2.03%) | 39.38 (-4.22%) | 54.06 (-3.32%) |
| FSS 0.4 | 27.97 (-72%) | 4264 (-70%) | 4.02 (+158%) | 53.92 (-4.75%) | 37.29 (-9.32%) | 52.89 (-5.40%) |
| FSS 0.5 | 16.28 (-84%) | 2540 (-82%) | 5.64 (+262%) | 51.55 (-8.94%) | 34.17 (-16.90%) | 50.92 (-8.93%) |
*PDE is the abbreviation of Percentage of Detections with Extraction. **FET is the abbreviation of Total Feature Extraction Time.
-
Download MOT17, MOT20 and DanceTrack datasets and organize them as follows: (Note: Place the DanceTrack val set in the train set, and the test set in the test set.)
data ├── MOT17 ├── test └── train └── MOT20 ├── test └── train └── dancetrack ├── test └── train -
Download prepared detections and features for MOT17 and MOT20, which are provided by StrongSORT from data in Google drive (or baidu disk with code "sort"). Place also weights for feature extractor (DukeMTMC_BoT-S50.pth) and place it under precomputed/.
Download the detections and features for DanceTrack, which are extracted using weights shared by OC-SORT from data in Google drive.
precomputed
├── MOT17_test_YOLOX+BoT # detections + features
├── MOT17_val_YOLOX+BoT # detections + features
├── dancetrack_test_YOLOX+BoT # detections + features
├── dancetrack_val_YOLOX+BoT # detections + features
├── MOT20_test_YOLOX+BoT # detections + features
├── DukeMTMC_BoT-S50.pth
- Set the paths of your dataset and other files in "opts.py", i.e., root_dataset dir_dets etc. The default paths are data/MOT17, data/MOT20, data/dancetrack, precomputed, etc.
A virtual environment is recommended to run the code. The code is tested on Python 3.8.19
pip install -r requirements.txtBase StrongSORT
python strong_sort.py {dataset} {dataset_type} --BoT --ECC --NSA --EMA --MC --woCFast-StrongSORT
python strong_sort.py {dataset} {dataset_type} --BoT --ECC --NSA --EMA --MC --woC --{offline} --{display} --ot {IoU_threshold} --aiou-threshold {ARS_threshold} --{decay_ema}Where offline indicates using precomputed features, display indicates saving the tracking results into the path indicated in opts.py
--ot, --aiou-threshold, --decay_ema are the parameters that are introduced in Fast-StrongSORT, and explained in the paper.
python strong_sort.py MOT17 val --BoT --ECC --NSA --EMA --MC --woC --offline --display --ot 0.2 --aiou-threshold 0.6 --decay_ema
python strong_sort.py MOT20 test --BoT --ECC --NSA --EMA --MC --woC --offline --display --ot 0.2 --aiou-threshold 0.6 --decay_ema
python strong_sort.py dancetrack val --BoT --ECC --NSA --EMA --MC --woC --offline --display --ot 0.2 --aiou-threshold 0.6 --decay_emaUse TrackEval. If the scores are too low, you probably have not checked out the following issue: dyhBUPT/StrongSORT#74