why set image_width = 1280 and image_height = 720?

[nianfd]

Great jobs!!
The image in dataset is 19201080, why set image_width = 1280 and image_height = 720 in this code?
I tried it and the effect was very poor when using other resolutions (e.g., when I set to 224224, the accuracy is 0). Can you tell me which parameters need to be modified if using other resolutions？

[dk-kim]

Thank you for your interest.
Our evaluation code assumes that the input resolution is 1280x720 (you can find in gt_tracks_24.txt file), so I think all of your box localization fails if you use different resolution.
Moreover, even if that's not the case, if the input resolution is too small, the feature map resolution will also be too low, making it difficult to extract proper human features. So, in my opinion, achieving good performance would be quite challenging.

[remember9069]

Thank you for your interest. Our evaluation code assumes that the input resolution is 1280x720 (you can find in gt_tracks_24.txt file), so I think all of your box localization fails if you use different resolution. Moreover, even if that's not the case, if the input resolution is too small, the feature map resolution will also be too low, making it difficult to extract proper human features. So, in my opinion, achieving good performance would be quite challenging.

Hello, and thank you for your outstanding work. I have a question related to this issue, specifically concerning this part

CAFE_codebase/models/models.py

Line 26 in caee21a

self.roi_align = RoIAlign(crop_height=self.crop_size, crop_width=self.crop_size)

I notice that spatial_scale and sampling_ratio are positional arguments.

Could you please clarify how the original import statement (from roi_align.roi_align import RoIAlign) handles these parameters without explicitly specifying them?

Thank you for your assistance.

[dk-kim]

Hi,
The RoIAlign code we used is from https://github.com/longcw/RoIAlign.pytorch, and as far as I know, it does not include spatial_scale and sampling_ratio as positional arguments.