search

lyuwenyu / RT-DETR

[CVPR 2024] Official RT-DETR (RTDETR paddle pytorch), Real-Time DEtection TRansformer, DETRs Beat YOLOs on Real-time Object Detection. 🔥 🔥 🔥
Apache License 2.0
2.62k stars 303 forks source link

Training on high resolution images #127

Open jarnokorstanje opened 1 year ago

jarnokorstanje commented 1 year ago

I appreciate your work a lot, thank you!

On the pytorch code, I am trying to train on a custom dataset of high-resolution images (8280x6208). In order to speed up the training process I have made a new dataset of positive tiles (1000x1000), which I am training on. When testing on these tiles, it performs very well!

Now, I tried testing on the full-size images with the trained weights, but performance is as low as it can possibly get. What would be the correct approach to train and test on high-resolution images?

lyuwenyu commented 1 year ago
  1. Do you modify L43 and L58 accoding your input size
  2. Make sure test input image resize with keep ratio

https://github.com/lyuwenyu/RT-DETR/tree/main/rtdetr_pytorch/configs/rtdetr/include#L43-L58

Pealse star this repo to follow updates