custom_data_processing.md

Processing your own data

Processing data for feature splatting is similar to Gaussian splatting. Currently, we only support colmap format data.

Colmap

To use colmap, first make sure colmap is available on your machine by running colmap -h and checking that help message is printed correctly. If not, please follow the instructions here.

Note: colmap works with or without CUDA. CUDA is automatically detected in colmap installation process by checking if nvcc is available. However, sometimes colmap can run into trouble with CUDA installed in conda. In this case, please run conda deactivate and return to the base environment before compiling colmap.

With installed colmap, let's say we create a sample_dataset under the feat_data folder. The user can put images into the input folder under the sample_dataset (i.e., feat_data/sample_data/input). Then we can run,

python convert.py -s feat_data/sample_data

to estimate information such as camera intrinsics/extrinsics and SfM sparse point clouds that help with Gaussian training.

Then we can compute features via

python compute_obj_part_feature.py -s feat_data/sample_data