finetuning results are much worse than general results, why??

[hengfei-wang]

There are other people reporting the same problem on DTU datasets and LLFF datasets. I have the same problem on my customized LLFF dataset. Can author give an answer about this problem?

[hengfei-wang]

I found that results get better if I delete --use_disp in

if is_finetued: cmd += f'--ckpt ./runs_fine_tuning/{scene}-ft/ckpts/latest.tar --use_disp'

The code is in renderer_video.ipynb file.

So the problem is still the difference between finetuning parameters and rendering parameters. I think author should mention users should check at lease use_color_volume, pad, and use_disp three setting in the code more explicitly.

[JJJano]

I think the problem lies in the difference parameters between fine-tuning and general, because my fine-tuning visualization results in DTU are not ideal.

[ch-ho00]

you can try changing pairs.th in the config file. the reference + source views of pairs.th aren't necessarily close as it is in the generalization training

[YuhsiHu]

Hi, thank you for sharing! The training dataset is like: Cameras, Depths and Rectified. How to transfer DTU dataset in a scanX format to fine-tune?