sunghwanhong / cost-aggregation-transformers Goto Github PK

Hi Sunghwan Hong,
Thanks for sharing the code. I have one question:

Will this be a potential bug at line https://github.com/SunghwanHong/Cost-Aggregation-transformers/blob/main/utils_training/evaluation.py#L32?

Should it be 'trg_kps' instead of src_kps?

Thanks

I found that this code use the target key points to prediction source keys points as shown in the code: https://github.com/SunghwanHong/Cost-Aggregation-transformers/blob/main/utils_training/evaluation.py#L25.

However, the other papers leverage the source key points to predict target points.

https://github.com/juhongm999/chm/blob/main/common/evaluation.py#L13

https://github.com/wintersun661/MMNet/blob/main/evaluation_tools/evaluation.py#L10

https://github.com/wookiekim/transformatcher/blob/main/logger.py#L179

Why not just follow the most common settings?

Hi thanks for the great work!

Just wondering when do you plan to release the pre-trained weights? It seems that link cannot be open.

Thanks!

Hi, thanks for the great work!

When I run the train.py, the data did not downloaded.
I got tarfile.ReadError: not a gzip file error.

Can you give dataset links that we can download manually?

Thank you!

Hi, thanks for the great work!

I think you have an error in the PCK-threshold of PF-Willow.
It is computed as the difference between the maximum and minimum kp. However, since you always pad the kp with -1, the minimum is always -1, which is not the actual coordinate of the minimum keypoint. Therefore, the pckthreshold is artificially large and therefore the results artificially high.
I know this error is present in multiple works, for example also in DHPF. However, the comparison is not fair to methods that actually use the correct metric, like NC-Net.

Hi,
I am trying to understand your code. You code is very clean and well arranged I have to admit. I have a 2 questions when reading your code.

1. May I ask why you pick hyperpixel_ids=[0,8,20,21,26,28,29,30]?
2. Have you tried other backbone other than resnet101?
   By the way, do you have any advices for me to improve your code? I tried to use resnet152 as backbone, but it does not perform better than resnet101. I guess it because I used the un-optimal hyperpixel_ids.

Thank you very much.

Hi, I trained the network on SPair-71k, and it seemed that it takes less 100 epochs to get fine result. So, how much epochs are trianed in the paper?

Hi,

As mentioned in issue 7, i would like to see the semantic correspondence matching between any of the image pairs in the test dataset.

Will you update the github with that script ?

In your training and testing code, the target image's points are transfered into source image, which is different to the setting in DHPF and CHM.
Is there anything wrong?

Is it possible to take the inference of your pretrained model on any novel image ? if yes, please provide the inference code ?

Hi,Thanks for your great work. I want to see the Visualization results of it. How can I draw images of correspondence between 2 images like Figure 5 in your paper. Do you have code for that? Thanks!

Hi,
Thanks for sharing your great work.
Would you mind adding a licence so that we will be more comfortable using your code?

Hi, thank you for your excellent work!
When will the CAT++ be open sourced?