Point cloud denoising library for Python.

Download the pixi package manager, this will manage the installation of all required packages.

Clone and install dependencies:

git clone https://github.com/n1n1n1q/pcd
cd pcd
pixi install

To run different examples:

pixi run {example-fft, example-ls, example-split, example-blob}

To activate conda enviroment:

pixi shell

The main focus of the project is exploring classical LA approaches for reducing noise in point clouds.

The FFT denoising method projects the point cloud onto a 2D grid, applies a Fourier transform to filter out high-frequency noise components, and then reconstructs the 3D point cloud from the filtered grid. This approach is particularly effective for point clouds with structured noise patterns.

The LS denoising method fits a local parametric surface to neighborhoods of points using least squares regression. By approximating the underlying smooth surface, this method can effectively remove random noise while preserving geometric features.

Local denoising applies denoising algorithms to subsets of the point cloud and then combines the results:

Local LS: Applies the least squares fitting algorithm to local patches of the point cloud. Local FFT: Applies Fourier filtering to local regions of the point cloud. Both local methods work by:

• Partitioning the point cloud into overlapping regions
• Processing each region independently using either LS or FFT
• Merging the denoised regions to form the final point cloud

To evaluate denoising performance, we use:

Chamfer Distance: Measures the average distance between points in the denoised cloud and their nearest neighbors in the original cloud Hausdorff Distance: Represents the maximum distance between a point in the denoised cloud and its nearest neighbor in the original cloud Processing Time: Records the computational efficiency of each method

For now, we have tested (global) FFT and local LS approaches. The followin results were obtained:

See visual comparison: Blob denoising comparison

Sphere denoising comparison

Although the project is almost finished, there is some work to do.

• Tweak FFT denoise
• Prepare a dataset with more models and upload it to GitHub
• Evaluate all the approaches (both global and local)
• Fix some misc functions (save, load etc)

• Basystyi Oleh
• Stasyshyn Anna
• Zhuk Maksym