ma-xu / Context-Cluster

[ICLR 2023 Oral] Image as Set of Points

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Model Question

Joazs opened this issue · comments

Hello, what is the effect of context cluster on small target segmentation? Is it suitable for small target segmentation? When I segmented small targets, I found that the effect was average. How can I improve it?

commented

Hi @Joazs, Sorry for the late reply since I was busy previously.

For small targets, I think you may need more centers/clusters, simplying update the proposal_w and proposal_h with greater values, and fine-tuning the model. More centers will focus more tiny objects, but will more computations.

I still don't understand. If more clustering centers are needed, it is more important to reduce the values of proposals_ w and proposal_ h?

commented

@Joazs we consider clustering centers = proposals_w * proposal_h. Help this can help. Feel free to let me know if you have any further questions.

commented

Closed no further discussions. Feel free to reopen it if any further questions.