Kneron doesn't support a lot of operations and new activation functions as mish, so we had to change mish to leaky ReLU in Scaled-YOLOv4-CSP and trained this model from scratch. Our custom model achieves with 448x448 image size 43.1% mAP @IoU=0.5:0.95 and 61.6% mAP @IoU=0.5 on COCO val 2017. Scaled-YOLOv4 repository: https://github.com/WongKinYiu/ScaledYOLOv4/tree/yolov4-csp
You can see difference between our model and suggested one in the table below. Fps calculated without postprocessing on the host side.
| model | mAP @ IoU=0.5:0.95 |
mAP @ IoU=0.5 |
FPS on KL-720 |
|---|---|---|---|
| Scaled-YOLOv4-CSP-leaky-448(ours) | 0.432 | 0.615 | 10.08 |
| pjreddie's YOLOv3-416 | 0.31 | 0.553 | 10.01 |
Model's accuracy after convertation and quantization slightly drops due to quantization. YOLOv3 showed here is the model that was converted in previous example https://github.com/SashaAlderson/Kneron_yolov3_inference.
| our model | mAP @ IoU=0.5:0.95 |
mAP @ IoU=0.5 |
|---|---|---|
| YOLOv3-416 | 0.252 | 0.489 |
| Scaled-YOLOv4-CSP-leaky-448 | 0.395 | 0.577 |
YOLOv4 is more optimal choise for Kneron KL-720 than YOLOv3. Also we used 500 images for quantization instead of suggested 10 in example, so our model loses significantly less accuracy(≈9%) comparing with YOLOv3(≈19%).