search

zeng-yifei / STAG4D

Official Implementation for STAG4D: Spatial-Temporal Anchored Generative 4D Gaussians
139 stars 6 forks source link

GT rendered images for evaluation #10

Closed ymxie97 closed 3 months ago

ymxie97 commented 3 months ago

Hi, Thanks for the great work!

The evaluation views on the dataset from Consistent4D should be different. Could you please provide the ground truth rendered images for evaluation?

Thanks!

zeng-yifei commented 3 months ago

The render settings we use for the evaluation is actually the same as Consistent4D. We provide the result of ours here if you need. stag4d.zip

ymxie97 commented 3 months ago

Thanks for your reply! So the evaluated images are after optimization (stage 2) instead of stage 1's output. Another question is why the Consistent4D's LPIPS reported in Table 1 is different from the numbers reported in Consisten4D paper if the evaluation setting is the same?

zeng-yifei commented 3 months ago

The view setting is the same, but the frames are not. In our setting, we use 30 frames to run for evaluation, so we rerun the eval script. Consistent4D gets a better LPIPS in this turn, and we don't know the reason too (laugh). However, we just put the result we tested in this table. That's why Consistent4D's LPIPS is better than the original paper.

ymxie97 commented 3 months ago

Thanks for your reply! Very appreciate it