search

yingkunwu / R-YOLOv4

This is a PyTorch-based R-YOLOv4 implementation which combines YOLOv4 model and loss function from R3Det for arbitrary oriented object detection.
114 stars 20 forks source link

How does it work with m x n input image size ? #61

Closed naga24 closed 10 months ago

naga24 commented 1 year ago

I have 8k images, what should i pass in the img_size param of train.py ?

yingkunwu commented 1 year ago

The img_size parameter determines the size to which images will be resized. You have flexibility in specifying the size, as long as it's divisible by 32. However, given the large size of your images, it would be better to split your images into several images with smaller size to achieve better results from the model.

naga24 commented 12 months ago

Yes, but how can i split the boxes corresponding the splitted images preserving the angle ?

yingkunwu commented 10 months ago

Sorry for the late reply. You can check out https://github.com/open-mmlab/mmrotate/blob/main/tools/data/dota/README.md to see how does that work. I used their code to generate split images for training as well. Hope this helps!

naga24 commented 10 months ago

Yes thanks for that I did the same from Dota dev kit

On Wed, 29 Nov, 2023, 3:46 am Ying-Kun Wu, @.***> wrote:

Sorry for the late reply. You can check out https://github.com/open-mmlab/mmrotate/blob/main/tools/data/dota/README.md to see how does that work. I used their code to generate split images for training as well. Hope this helps!

— Reply to this email directly, view it on GitHub https://github.com/kunnnnethan/R-YOLOv4/issues/61#issuecomment-1830846780, or unsubscribe https://github.com/notifications/unsubscribe-auth/AC4ED6XXGZJN4Y6B2VXI3SLYGZPFBAVCNFSM6AAAAAA6CAAAOWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQMZQHA2DMNZYGA . You are receiving this because you authored the thread.Message ID: @.***>

yingkunwu commented 10 months ago

I have noticed that there are some bugs in the DOTA dev kit. I could not split ground truths according to images properly using DOTA dev kit. I would recommend you to validate it. If it works for you then never mind.