facebookresearch / PoseDiffusion

[ICCV 2023] PoseDiffusion: Solving Pose Estimation via Diffusion-aided Bundle Adjustment
Other
718 stars 42 forks source link

Guidance Sampling or Direct Optimization? #40

Open gitouni opened 3 months ago

gitouni commented 3 months ago

Thanks for sharing the code. I have a question about the implementation of Sampson-guided sampling. As illustrated in Eq (8) in your paper, the probability of p(I|x_t) is a scaled gradient imposed on the noise predicted by the model, which is consistent with this paper. However, for corresponding implementation I found here: https://github.com/facebookresearch/PoseDiffusion/blob/36eeb1654ad8e67844672d7a40e7c179e9c58104/pose_diffusion/util/geometry_guided_sampling.py#L67-L126

https://github.com/facebookresearch/PoseDiffusion/blob/36eeb1654ad8e67844672d7a40e7c179e9c58104/pose_diffusion/util/geometry_guided_sampling.py#L129-L172

It seems that the model_mean is directly optimized by sampson-error rather than placed in the loop of a standard guidance sampling.

I apologize if I misunderstood your code. Appreciate your reply.