From the perspective of diffusion model principle, why are the representations generated by DDPM from remote sensing images more robust and distinguishable than those obtained by UNet networks?

wgcban / ddpm-cd

Remote Sensing Change Detection using Denoising Diffusion Probabilistic Models

https://www.wgcban.com/research#h.ar24vwqlm021

MIT License

264 stars 32 forks source link

From the perspective of diffusion model principle, why are the representations generated by DDPM from remote sensing images more robust and distinguishable than those obtained by UNet networks? #44

Open ZhiyuLong0328 opened 5 months ago

ZhiyuLong0328 commented 5 months ago

Hello. Your article DDPM-CD has been very helpful to me, and I have a question I would like to discuss with you. The question is as follow.

From the perspective of diffusion model principle, why are the representations generated by DDPM from remote sensing images more robust and distinguishable than those obtained by UNet networks?

Looking forward to your reply, thank you. Sincerely yours.

jehovahxu commented 3 months ago

I don't understand either. When T is large enough, the feature is approximate to gaussian noise. It is very different to use an auto-encoder such as UNet directly. So it is hard to understand why the image representations can be extracted from gaussion noise in the diffusion process.

ZhiyuLong0328 commented 2 months ago

Can you understand why multi-time step features are more effective for change detection tasks than single-step feature? What additional information can it extract, and why are multiple time steps mutually reinforcing?