Data process

[Simon4Yan]

Thank you for your nice work! I might missed something so I met an issue with the data process:

First case: I use --data_format colmap to load my own data (colmap):

Second case: I first run your torch-ngp python scripts/colmap2nerf.py to generate transform_train/val/test.json. Then, I use --data_format nerf to load my own data.

The two cases lead to different results: the first one gets 18 in PSNR while the latter produces 22: the visualizations also reflect such a performance gap.

May I know your suggestion on this? Thanks.