search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.79k stars 7.97k forks source link

608 input size for yolov4-tiny #8011

Open forestguan opened 3 years ago

forestguan commented 3 years ago

I directly use yolov4-tiny.weights and change width/height to 608 in yolov4-tiny.cfg, then I use this config to validate performance on the COCO testdev. The result of AP50 is only 0.37. Is that correct? May be I should re-training the model under 608 input size ? Sincerely

Duarte-Nunes commented 3 years ago

Hello, if you want the network to work with 608x608 on input you have to train it on that input size. You cant just take weights trained on another size and change the cfg.

stephanecharette commented 3 years ago

You cant just take weights trained on another size and change the cfg.

You definitely can change the size in the .cfg afterwards to make it work. This is well-documented. But I'm not surprised that the results aren't quite as good.