| Dataset | Faster-RCNN | mAP@50 | YOLOv8 | mAP@50 |
|---|---|---|---|---|
| VG-150 | Download | 28.10 | yolov8m_vg150.pt | 26.48 |
| IndoorVG | Download | 25.31 | yolov8m_indoorvg.pt | 36.65 |
| PSG | Download | 35.38 | yolov8m_psg.pt | 53.60 |
| yolov8x_psg.pt | 57.20 | |||
| yolov9m_PSG | 51.60 | |||
| yolov10m_PSG | 51.20 |
All models with YoloV8 and YoloV8-World are trained without any debiasing or re-weighting methods (such as TDE or reweight loss) and performance could probably be further improved.
New weights for the REACT model with YOLOV8-m backbone (SGDET only):
| Models | WEIGHTS | R@20 | R@50 | R@100 | mR@20 | mR@50 | mR@100 | mAP | Latency (ms) |
|---|---|---|---|---|---|---|---|---|---|
| REACT (YOLOV8m) | Download | 28.46 | 31.88 | 33.37 | 17.66 | 19.81 | 21.89 | 53.57 | 23.9 |
New weights for the REACT model with YOLOV8-m backbone (SGDET only):
| Models | WEIGHTS | R@20 | R@50 | R@100 | mR@20 | mR@50 | mR@100 | mAP | Latency (ms) |
|---|---|---|---|---|---|---|---|---|---|
| REACT (YOLOV8m) | Download | 21.04 | 26.16 | 28.75 | 9.78 | 12.26 | 13.63 | 31.8 | 23.9 |
Please download the weights in a checkpoints folder at the root of the codebase and run visualization using the demo/SGDET_on_cutom_images.ipynb notebook or evaluation using tools/relation_test_net.py.