Kin-Yiu, Wong

In this case, there are 4 pyramid outputs, 3 anchors for each output. so anchors is 3.

You may need modify class loss part, and maybe also target generation part.

You have to use train.py instead.

Mainly because of paste_in I think, currently we implement paste_in by a for loop, it very slow when CPU is not powerful.

For keypoint detection: merge [this](https://github.com/WongKinYiu/yolor/releases/download/weights/yolor-pose.zip) into [this](https://github.com/TexasInstruments/edgeai-yolov5/tree/yolo-pose) and modify [this](https://github.com/WongKinYiu/yolov7/blob/main/cfg/training/yolov7-w6.yaml) for training. You could update your code on it, I just add YOLOR support.

SimOTA uses top K, OTA uses optimization.

I think you are correct. Thanks for point out of this.

I guess your are running batch 32 inference. For batch 32 inference, YOLOv7 takes 2.8 ms average inference time, and YOLOv5m takes 1.7 ms average inference time in the paper.

What are the inference time you get on yolov7-tiny, yolov7, yolov5n, yolov5s, yolov5m, and yolov5l.

CPU inference time is usually proportional to FLOPs.