search

WongKinYiu / yolov7

Implementation of paper - YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors
GNU General Public License v3.0
13.31k stars 4.2k forks source link

"Training from scratch" #221

Open StarBurstStream0 opened 2 years ago

StarBurstStream0 commented 2 years ago

Hi, I wonder if YOLOv7 will keep supporting "Training from scratch" strategy, and would not provide imagenet pertained weights in the future? If so, Im very interested in the training process about how YOLOv7 succeed other imagenet pertained models on performance? Which are the key factors?

WongKinYiu commented 2 years ago

Hmm... We have some experiments on different backbones in timm. In actually, ImageNet pretrained still help coco object detection task. If ImageNet22k pretrained model is used, it may get +2~4% AP then model which is trained from scratch.