Here we provide the list of our current available models and their performance on various benchmarks. For TAO TETA, we use public detections from Detic-SwinB. For BDD MOT, we use public detections from ByteTrack, Yolo-X. For BDD MOTS, we use public detections from UNINEXT-H.
| TAO TETA | BDD MOT | BDD MOTS | model | ||||
|---|---|---|---|---|---|---|---|
| AssocA | TETA | AssocA | mIDF1 | AssocA | mIDF1 | ||
| TETer | 36.7 | 34.6 | 52.9 | 51.6 | - | - | - |
| OVTrack | 36.7 | 34.7 | - | - | - | - | - |
| ByteTrack | - | - | 51.5 | 54.8 | - | - | - |
| UNINEXT | - | - | - | 56.7 | 53.2 | 48.5 | - |
| MASA-R50 | 43.0 | 45.8 | 52.1 | 55.1 | 54.8 | 50.2 | HF 🤗 |
| MASA-Sam-B | 44.3 | 46.5 | 53.1 | 55.6 | 54.1 | 49.7 | HF 🤗 |
| MASA-Sam-H | 44.6 | 46.4 | 52.4 | 55.2 | 54.8 | 50.1 | HF 🤗 |
| MASA-Detic | 44.1 | 46.3 | 53.3 | 55.8 | 53.8 | 49.9 | HF 🤗 |
| MASA-GroundingDINO | 44.6 | 46.7 | 53.1 | 55.7 | 53.3 | 48.9 | HF 🤗 |
- Note that during MASA training, we do not use any in-domain images. The results are slightly higher than we reported in the paper.
For more details, please refer to the benchmark_test.md file.