by Hye-Jin Park, Young-Ju Choi, Young-Woon Lee, Byung-Gyu Kim
This is an unofficial implementation of “ssFPN: Scale Sequence (S^2) Feature Based-Feature Pyramid Network for Object Detection”
The implementation was only trained and tested on my private dataset due to resource and time issues
My private dataset contains 8w images taken with real-world 1080P cameras, totaling 26 categories.
PS: MultiScale Training is not used on ssFPN because of training memory issues.
| Model | pre-train | epochs | mAP.5 | mAP.5:.95 |
|---|---|---|---|---|
| Baseline(YoloV5 6.0) | yolov5s.pt | 150 | 0.901 | 0.655 |
| Baseline + SpaceToDepth | yolov5s.pt | 150 | 0.901 | 0.657 |
| Baseline + SpaceToDepth + ssFPN | yolov5s.pt | 150 | 0.909 | 0.674 |
| Model | resolution ratio | GPU | FP16 latency(ms) |
|---|---|---|---|
| Baseline(YoloV5 6.0) | 384x640 | MX450 | 15-21 |
| Baseline + SpaceToDepth | 384x640 | MX450 | 15-18 |
| Baseline + SpaceToDepth + ssFPN | 384x640 | MX450 | 27-30 |
Special thanks to w13ww for providing the coco2017 test results
| Model | pre-train | epochs | mAP.5 | mAP.5:.95 | AP Small | AP Medium | AP Large |
|---|---|---|---|---|---|---|---|
| Baseline(YoloV5 6.0) | yolov5s.pt | 150 | 0.574 | 0.376 | 0.217 | 0.423 | 0.492 |
| Baseline + ssFPN | yolov5s.pt | 150 | 0.579(↑0.005) | 0.380(↑0.004) | 0.234(↑0.017) | 0.430(↑0.007) | 0.478(↓0.014) |
| Model | GPU | FP16 latency(ms) |
|---|---|---|
| Baseline(YoloV5 6.0) | A100 | 0.9 |
| Baseline + ssFPN | A100 | 1.3 |
