Testing Pretrained Models

[ChenYutongTHU]

Hi authors,

I have an issue of evaluating the provided pretrained models. Following https://github.com/zju3dv/animatable_nerf/issues/63, I downloaded the vanilla animatable NeRF models， aligned_aninerf_lbw_${sub}, and run the command for test, according to https://github.com/zju3dv/animatable_nerf/blob/master/test.sh

# Vanilla Animatable NeRF
python run.py --type evaluate --cfg_file configs/aligned_nerf_lbw/aligned_aninerf_lbw_${sub}.yaml exp_name aligned_aninerf_lbw_${sub} resume True
python run.py --type evaluate --cfg_file configs/aligned_nerf_lbw/aligned_aninerf_lbw_${sub}.yaml exp_name aligned_aninerf_lbw_${sub}_full resume True aninerf_animation True init_aninerf aligned_aninerf_lbw_${sub} test_novel_pose True

However, the resulted metrics are somehow different from the reported ones in NeuralBody-TPAMI.

For example, for subject 393 (Swing3), the reported novel-view scores are PSNR=27.53, SSIM=0.925, the yielded scores of the downloaded model, aligned_aninerf_lbw_393/latest.pth are PSNR=25.58, SSIM=0.9121

Is there any reason for this? (For example, the dataset was updated)

[ChenYutongTHU]

I see aligned_aninerf_lbw_${sub}.yaml uses data/light_stage while configs/aninerf_${sub}.yaml uses data/zju_mocap. Is there any difference between light_stage and zju_mocap?

[pengsida]

There is no difference between two data.

[ChenYutongTHU]

Thanks! Are rendered images with the reported scores available? If the released model cannot reproduce the reported scores on my machine, it would be great to access the rendered results directly.

Thank you :)

[ChenYutongTHU]

Also, can you help provide the init_sdf model, which seems to be needed in the Ani-SDF method. Great thanks!