Awesome
A Library of Multi-Object Tracking in Python and Pytorch
Installation
environments: python 3.6.10, opencv 4.1.1, pytorch 1.3+
git clone https://github.com/nightmaredimple/libmot --recursive
cd libmot/
python setup.py install develop --user
The details can be seen from my blogs or zhihu.
Feature Lists
| Block | Method | Reference | Complete |
|---|---|---|---|
| IOU Assignment | iou-tracker&V-IOU | ✓ | |
| Linear Assignment | - | ✓ | |
| Data Association | MinCostFlow | MCF | ✓ |
| Other End-to-End Network | DAN&DeepMOT | ☐ | |
| GNN&GCN | MPNTrack | ☐ | |
| ----------------------- | ----------------------------------- | ---------------------- | --- |
| Kalman Filter | Sort&DeepSort | ✓ | |
| Motion | ECC | Tracktor++ | ✓ |
| Epipolar Geometry | TNT | ✓ | |
| ----------------------- | ----------------------------------- | ---------------------- | --- |
| Re-ID | - | ☐ | |
| Appearance | Feature Fusion&Selection | - | ☐ |
| DAN | DAN | ✓ | |
| ----------------------- | ----------------------------------- | ---------------------- | --- |
| Detection | Faster RCNN + FPN | Tracktor++ | ☐ |
| ----------------------- | ----------------------------------- | ---------------------- | --- |
| SOT | CF&Siam | KCF&CN | ☐ |
| ----------------------- | ----------------------------------- | ---------------------- | --- |
| DataLoader | - | ✓ | |
| Tricks | Spatial Blocking | - | ✓ |
| ----------------------- | ----------------------------------- | ---------------------- | --- |
| Evaluation | - | ✓ | |
| Others | Tracking Visualiztion | - | ✓ |
| Feature Visualiztion | - | ☐ | |
| ----------------------- | ----------------------------------- | ---------------------- | --- |
| Tracktor | MIFT(ours) | - | ☐ |
| ----------------------- | ----------------------------------- | ---------------------- | --- |
| Detector | MIFD(ours) | - | ☐ |
Motion Model
python scripts/test_kalman_tracker.py
Data Association
<div align="center"> <img src="figures/linear_assignment.png" /> </div>Tracktor
Our proposed MIFT and MIFD will be released upon acceptance.
In MOT Challenge, the MIFT tracktor is named as ISE-MOT, the MIFD detector is named as ISE-MOTDet.
| Method | DataSets | MOTA↑ | IDF1↑ | MT↑ | ML↓ | FP↓ | FN↓ | ID Sw.↓ | Frag↓ | Hz↑ |
|---|---|---|---|---|---|---|---|---|---|---|
| MOT15 | 48.1 | 52.1 | 29.5% | 26.2% | 10246 | 20840 | 776 | 1197 | 6.7 | |
| MIFT | MOT16 | 60.4 | 57.3 | 24.6% | 28.9% | 5510 | 66723 | 704 | 932 | 6.9 |
| MOT17 | 60.1 | 56.2 | 28.1% | 27.8% | 22265 | 200077 | 2644 | 3206 | 7.2 |
| Method | DataSets | AP↑ | MODA↑ | FAF↓ | Precision↑ | Recall↑ |
|---|---|---|---|---|---|---|
| MIFD | MOT17Det | 0.88 | 70.7 | 4.1 | 81.4 | 91.7 |
To be continued..