search

justinkay / aldi

Official implementation of "Align and Distill: Unifying and Improving Domain Adaptive Object Detection"
https://aldi-daod.github.io/
47 stars 7 forks source link

How to train UMT with this project ? #31

Closed BalajiVenkatesan006 closed 2 months ago

BalajiVenkatesan006 commented 2 months ago

Is there any documentation of configs to replicate all these mentioned experiments in paper? on custom dataset?

BalajiVenkatesan006 commented 2 months ago

I am using the extras branch for training the UMT on my custom dataset, what is the first step for this? How to train Cycle gan to get the fake images that I need to give as a part of input in the UMT training. as specified in the dataset.

justinkay commented 2 months ago

Hi @BalajiVenkatesan006, for a custom dataset you'll need to train your own CycleGAN with this codebase: https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix and then add make sure your register_coco_instances call includes a dictionary with the field translated_image_dir that points to where your generated files are. See: https://github.com/justinkay/aldi/blob/a946600c7a95b1bd01d7e5dd23e3e8ab94f3fe03/aldi/datasets.py#L22

We have some details of how we trained the CycleGAN in the Supplemental section 3. I'll let @timmh chime in too in case there are any additional details you should know.

BalajiVenkatesan006 commented 2 months ago

Thanks for your instant reply. It helped.

timmh commented 2 months ago

@BalajiVenkatesan006 the info in the supplemental should indeed suffice to reproduce our results but feel free to let me know in case you have additional questions.