search

fudan-zvg / 4d-gaussian-splatting

[ICLR 2024] Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting
MIT License
616 stars 45 forks source link

Poor results using custom data #47

Open henrypearce4D opened 2 months ago

henrypearce4D commented 2 months ago

I cant get any clean results from using custom data.

00000

I created a transforms.json by converting my colmap model for one frame using the colmap2nerf script from instant-ngp, this trains great in instant-ngp; https://github.com/NVlabs/instant-ngp/blob/master/scripts/colmap2nerf.py

Then created a script that duplicates the libraries for a specified amount of frames and adds a time stamp, here is my resulting .json, can you see any issues with it @Alexander0Yang ? does your code accept per camera intrinsics?

transforms_train.json

Alexander0Yang commented 2 months ago

Hi @henrypearce4D, I just checked your transforms.json, it seems no problem. Have you tested your date on the 3D Gaussian Splatting?

henrypearce4D commented 2 months ago

Hi @Alexander0Yang, I tested the transforms.json with instantNGP, I dont think the Inria code that can load .json / synthetic blender datasets, that have varied camera intrinsics. we have 5 camera models (varied pixel width and height and lenses) in this dataset

Can you recommend a way to test the .json with any standard gaussian splatting method?

ax-anoop commented 1 week ago

@henrypearce4D

I added this my self into the code base, so let me know if you still need help!