Verg-Avesta / CounTR

CounTR: Transformer-based Generalised Visual Counting
https://verg-avesta.github.io/CounTR_Webpage/
MIT License
92 stars 9 forks source link

question about density map #38

Closed zhitao654321 closed 9 months ago

zhitao654321 commented 10 months ago

Hello. Thank you for the great work. I got curious. Why did you choose to mark dots on the ground truth density map, instead of using original density map with Gaussian distribution from FSC147 dataset ? Just wondering about your thoughts on this. Thanks

Verg-Avesta commented 10 months ago

Hello. Thanks for your attention.

In the original density map of FSC-147, the writers used an adaptive window size for Gaussian smoothing, which means different objects have different size of annotations. Thus, the model have to learn object size information besides counting. To remove the effect of different object size, we re-generate the Gaussian density map with a fixed window size based on the point-annotations.

zhitao654321 commented 10 months ago

I get it, thank you very much