Some question about equation (2) in your paper

[Jiang-HB]

Thanks for your excellent and inspiring work. I have something confusing.

As we all know, after the diffusion process, the common diffusion model like DDPM would use the posterior distribution q(z_{t-1} | z_t, z_0) based on the bayesian theorem to guide the learning of prior distribution p(z_{t-1} | z_t). However, equation (2) in your paper shows you seem to directly use z_0 as the optimization target to guide the learning of prior distribution. In my opinion, such a difference leads to the proposed DiffusionDet just using the diffusion model to perform data augmentation, which raises a new question of "whether choosing another data-augmentation technique also can bring a similar performance?"

Sincerely hope to receive your reply.

[ShoufaChen]

Hi, Thank you for your feedback. I apologize for the delayed response due to the holiday season.

In our work, we employ the posterior distribution q(z_{t-1} | z_t, z_0) to train the prior distribution p(z_{t-1} | z_t). Our objective is to model the posterior distribution throughout the diffusion process. Our approach is different from data augmentation techniques in that it explicitly models the diffusion process.

[Tao-DoubleNine]

Thanks for your excellent and inspiring work, I also have a similar issue. Your loss function is related to z0 and the results of the model output, which is more like a process with only adding noise, because your network can directly predict z0, it seems unnecessary to infer and sample again.