Finetuned model weights

[vineet2104]

Hi,

Would it be possible to release the finetuned model weights of CROG and any inference scripts?

[ole2412]

@HilbertXu
Dear Sir,
thanks for open sourcing this amazing CROG.
We try to reproduce the results reported on the OCID-VLG dataset, however we miss the finetuned model weights used to produce the results.

Could you please give me some advice about how to fix this problem? Thank you in advance for your kindly help!

Regards,
Ole

[HilbertXu]

Hi @ole2412, thanks for your interest in out work.
Unfortunately, we dont have existing weights, I think I can one from remote server but cannot guarantee it is the lastest one. I will put a google drive link under this issue so that you can download it.

Regarding your another question about the missing "build_segmentor", I'm sorry for this mistake. This repo was rebuilt and used for our other project as a template (you can see that there is another model named SSG in model dir), thus some function names were changed. I think if you want to use CROG model you can change the "build_segmentor" to "build_crog".

Thanks,
Hilbert

[ole2412]

@HilbertXu
Thank you for your response!
Unfortunately, I can not see the drive link.
Where can I find it?

Thanks,
Ole

[HilbertXu]

@ole2412
I just found one from the remote server and currently testing the weight, just to make sure it is not a corrupted one. If the weight is usable, I will put a download link and @ you.

[HilbertXu]

@ole2412
Here is the download link: https://drive.google.com/file/d/14OyUHE9YvMRYJT-xm5mPq_bf_LmgE3k8/view?usp=sharing

I also updated the test script, you can run inference with CROG model by running the following command:

python test_crog.py --config config/OCID-VLG/crog_multiple_r50.yaml

Here is what I got with this checkpoint,

This checkpoint did not yields IoU results as reported in paper, but it could be used as a good start point for further fine-tuning.