search

zhizdev / sparsefusion

[CVPR 2023] SparseFusion: Distilling View-conditioned Diffusion for 3D Reconstruction
https://sparsefusion.github.io/
362 stars 14 forks source link

this metric is used for debugging only and not the final metric #8

Closed wzf2022 closed 1 year ago

wzf2022 commented 1 year ago

Hi, I've run demo.py without modifications, and the results are close to the figures mentioned in the paper. And one warning is that 'this metric is used for debugging only and not the final metric'. I wonder how can I get the final metric？Thanks!

zhizdev commented 1 year ago

Hi, thanks for running our demo!

In CO3D, the square cropped images may come with padded regions. The padded region may abruptly cut off an object (when the object is cut off in the original smartphone video capture). Therefore, we must mask our predicted results with the padded region properly to match the padding on the ground truth images. This valid region correspond to 'valid_region':frame_data.mask_crop in the CO3D dataloader.

zhizdev commented 1 year ago

Closing due to inactivity. Feel free to reopen.