1.1.0

Tested core features were extended with thesis-specific materials to replicate the study: experiment runner scripts, WandB logs, and visualization code in a form of jupyter notebook

1.0.0

Release of All Planned Features

Added

• MuSHRoom colmap cameras parsing
• Point cloud statistical denoising and optional random downsampling to fixed size
• Batch experiment runner script that integrated point cloud reconstruction and 3D Gaussian Splatting training into a single scene-wise experiment (and single corresponding Weights&Biases run) is now live, tested, and documented

Fixed

• reconstruction.skip_n is now a child-parameter of reconstruction and if it is > 1, unneded images, depth maps and cameras won't be loaded into RAM, meaning that less compute resources are used at initialization
• ARKit intrinsic matrices are parsed well now
• reconstruction.parameters.voxel_downsample_size is not ignored anymore
• Error handling is more versatile now

Figure 1. Example of WandB log for several launches of batch runner script

0.2.0

Added

• Documentation in README.md with setup, usage, and extension instruction. Also added short datasets documentation
• MuSHRoom dataset Parsers
• ICP Colored registration to reconstruction pipeline
• multiple new parameters in configs/main.yaml for reconstruction
• Fork of Gaussian Splatting with WandB logging as submodule
• requirements.txt for main pipeline

Figure 1. Example of ICP registration with camera pose refinement. Green points represent existing point cloud, red points are the new point cloud, blue points are from aligned new point cloud. Red and blue cameras are new camera and refined new camera after registration

0.1.0

MVP Release

Added:

• Base classes structure
• Hydra configuration
• RGBD conversion to point cloud from custom ARKit logger
• Reconstruction visualization