| ModelType | InitialWeight | mAP50 | config |
|---|---|---|---|
| v1-original | yolov3.weights | 88.8 | - |
| v2-original | darknet53.conv.74 | 83.3 | - |
| v3-original | mobilenetv2 | 78.9 | - |
| v1-this | yolov3.weights | 86.2 | strongerv1.yaml |
| v2-this | darknet53.conv.74 | 80.2 | strongerv2.yaml |
| v3-this | mobilenetv2 | 79.6 | strongerv3.yaml |
| v3-this | mobilenetv2-0.75 | 76.97 | strongerv3_0.75.yaml |
Note:
- This project use threshold=0.1 for faster evaluation,while the original implementation use 0.01.
- Adjust the training schedules(total epochs,lr scheduler) may further boost the performance. I pick StepLR instead of ConsinLR to accelerate training procedure. Continue training may give better results.
- data augmentation
- multi scale train
- focal loss
- soft nms
- mix up
- label smooth
- consine learning rate(Just replace the lr scheduler.)
- GIOU
- multi scale test (See another repo of mine for details.)