ultralytics / yolov3

YOLOv3 in PyTorch > ONNX > CoreML > TFLite
https://docs.ultralytics.com
GNU Affero General Public License v3.0
10.16k stars 3.44k forks source link

multi-scale must be with yolo-spp.cfg? #484

Closed ZhxJia closed 5 years ago

ZhxJia commented 5 years ago

thanks for your work.I have a question need your help. when the multi-scale set true ,whether the spp is necessary?

glenn-jocher commented 5 years ago

@ZhxJia spp is an architecture, multi-scale is a training technique. They are not related.

glenn-jocher commented 5 years ago

@ZhxJia so you can use --multi-scale with yolov3.cfg or yolov3-spp.cfg, they are independent of each other. Does this make sense?

ZhxJia commented 5 years ago

Thanks for your reply.I see. multi-scale means the output grid will be different,is my understanding correct?

glenn-jocher commented 5 years ago

@ZhxJia --multi-scale means that the training image sizes will vary +/- 50% about --img-size. It's a very common training technique.