NeRF synthetic + mipNeRF360 dataset training fails.

[bchao1]

I tried to run 3DGS on the nerf synthetic data (chair example) and mipNeRF360 dataset, but the training results were really bad. I used the following command, as suggested in the README (-w for white background):

For NeRF synthetic:
python train.py -s data/nerf_synthetic/chair/ --eval -w

For mipNeRF360:
python train.py -s ~/data/360_v2/bicycle/ -i images_4 --eval

The train/test PSNRs are around 10-15 dB. Did I incorrectly set any parameters? Thanks!

Hardware: NVIDIA RTX A6000

[Futuramistic]

I also saw that issue with the training! I thought it was an issue with my setup (CUDA 12.4) and spent the whole day setting it up to the older version, but it turned out it was somewhere in the repo. I reverted to commit 21301643a4354d6e24495c0df5a85354af8bd2be and I get the correct PSNR.

[alanvinx]

Hi, did you try to render the tests images after training and computing the metrics using metrics.py?
It looks like the values printed from the training report: https://github.com/graphdeco-inria/gaussian-splatting/blob/main/train.py#L158C13-L158C28 are not correct, this will be fixed soon.

[alanvinx]

Hi, this is now fixed (3dc0b75).

[Huster-YZY]

training_report(tb_writer, iteration, Ll1, loss, l1_loss, iter_start.elapsed_time(iter_end), testing_iterations, scene, render, {'pipe':pipe, 'bg_color':background, 'use_trained_exp': dataset.train_test_exp, 'separate_sh':SPARSE_ADAM_AVAILABLE}, dataset.train_test_exp) if (iteration in saving_iterations): may be more elegant :)