search

NVlabs / InstantSplat

InstantSplat: Sparse-view SfM-free Gaussian Splatting in Seconds
https://instantsplat.github.io/
Other
861 stars 53 forks source link

How can I use my own data to train and test? #16

Open gwxxx opened 2 months ago

Yenhongxuan commented 2 months ago

I have the same question. How can I train and evaluate on dataset like ScanNet? Also, how to prepare the ground truth pose in ScanNet so that we use the ground truth pose for evaluating pose estimation instead of using poses from COLMAP?

kairunwen commented 2 months ago

How can I use my own data to train and test?

Hi, we slightly modified the FSGS script to obtain the ground truth pose and other relevant data. Feel free to modify it for your own dataset.

Yenhongxuan commented 2 months ago

How can I use my own data to train and test?

Hi, we slightly modified the FSGS script to obtain the ground truth pose and other relevant data. Feel free to modify it for your own dataset.

Sorry, the script provided seems to retrieve poses and other data from the original COLMAP data if I don't misunderstand. However, it appears that only RGB images and images.bin (Used to evaluate the pose estimation) are needed. How could I generate the poses using ground truth poses from ScanNet but not retrieve them from COLMAP data?

Please correct me if there are any mistakes.

kairunwen commented 2 months ago

Sorry, the script provided seems to retrieve poses and other data from the original COLMAP data if I don't misunderstand. However, it appears that only RGB images and images.bin (Used to evaluate the pose estimation) are needed. How could I generate the poses using ground truth poses from ScanNet but not retrieve them from COLMAP data? Please correct me if there are any mistakes.

You can write a script to read ScanNet data and save it in COLMAP format. Reference： (1) https://github.com/colmap/colmap/blob/main/scripts/python/read_write_model.py (2) https://github.com/graphdeco-inria/gaussian-splatting/blob/main/scene/colmap_loader.py