What the `ddim_step` does in the Algorithm 2 of the paper

[GT9505]

Hi~, I wonder to know what the DDIM does in the ddim_step of Algorithm 2 of the paper. Could you describe intuitionally how the DDIM changes the pred_bboxes. I don't get why DDIM can improve the bbox AP.

[tyshiwo1]

...same question

[ShoufaChen]

Hi, thanks for your interests in our work.

DDIM [a] is a variant of DDPM [b] with the goal of accelerating sampling process while maintaining performance. It is used to generate the input of next sampling step from the output of current step. More specifically, DDIM is implemented here in our code.

Intuitively, DiffusionDet is trained with noisy boxes, while the output of sampling step is the prediction of the network, instead of 'noise'. Therefore, DDIM is used to transform the prediction to 'noisy distribution'.

[a] Song, Jiaming, Chenlin Meng, and Stefano Ermon. "Denoising diffusion implicit models." ICLR 2021.

[b] Ho, Jonathan, Ajay Jain, and Pieter Abbeel. "Denoising diffusion probabilistic models." NeurIPS 2020.

[tyshiwo1]

Thank you for your reply.

Hi, thanks for your interests in our work.

DDIM [a] is a variant of DDPM [b] with the goal of accelerating sampling process while maintaining performance. It is used to generate the input of next sampling step from the output of current step. More specifically, DDIM is implemented here in our code.

Intuitively, DiffusionDet is trained with noisy boxes, while the output of sampling step is the prediction of the network, instead of 'noise'. Therefore, DDIM is used to transform the prediction to 'noisy distribution'.

[a] Song, Jiaming, Chenlin Meng, and Stefano Ermon. "Denoising diffusion implicit models." ICLR 2021.

[b] Ho, Jonathan, Ajay Jain, and Pieter Abbeel. "Denoising diffusion probabilistic models." NeurIPS 2020.

[ShoufaChen]

Hi, @huilicici

Thanks for your interest. Would you mind opening a new issue to discuss your concern about the loss since it is out the scope of this issue.

[huilicici]

Hi, @huilicici

Thanks for your interest. Would you mind opening a new issue to discuss your concern about the loss since it is out the scope of this issue.

All right. Thank you.