Thesis: Processing of Unstructured Point Clouds using Neural Networks
Tutor: RNDr. Martin Madaras, PhD.
Neural networks have been found as a very efficient machine learning tool for image processing tasks.
The structured point clouds have the same structure as the 2D image data, with the additional 3D
position information, thus the neural networks designed for 2D image processing can be used for
these 3D data. Moreover, in the next step of the processing, these multi-view structured point
clouds are merged together into one unstructured point cloud. The unstructured point clouds cannot
be processed in the same way directly as the structured point cloud. The domain of the unstructured
point cloud needs to be transformed first, in order to be processed by a convolution neural network.
The goal of the thesis is to analyze existing neural network models that are used for the
processing of unstructured point clouds. Furthermore, take models of neural convolution networks
that are used for image processing and propose a network for processing of point clouds. Compare
proposed models with the existing ones and evaluate the results. Apply proposed models for
processing of point clouds that are obtained as results from 3D scanning. Interesting processing
steps might be intensity or color normalization, position smoothing or another local filtering.
- Study of materials on existing models of neural networks which process unstructured point clouds (15.11.2020)
- Test implementation of atleast two of the models, e.g. OctNet model & PointNet (24.11.2020)
- Presentation of gathered resources (01.12.2020)
- Analysis, comparison and possible reimplementation (08.12.2020)
- Implement voxelization of unstructured point clouds including one of studied methods (15.12.2020)
- Try to apply the implementation on 3D objects, unstructured point clouds, which are gathered as union of
structured 3D scans (scans from various points of view), analyse and gather results (19.01.2021)
- Try to solve some of the interesting processing steps on unstructured point clouds
like classification, artifact filtering, intensity/color normalization or another local filtering (09.03.2021)
- Conclusion and comparison between my implementation (voxelization) and other approaches (13.04.2021)
- Research on classification models (17.02.2021)
- Classification model implemented (24.02.2021)
- Research on voxel segmentation models (03.03.2021)
- Figure out voxel segmentation pipeline (10.03.2021)
- Implementing voxel based segmentation model (17.03.2021)
- Voxel based segmentation implemented (24.03.2021)
- Code refactoring, making results (31.03.2021)
- Classification on real life scans, paper writing (07.04.2021)
- Better reversed mapping for segmentation results, paper writing (14.04.2021)
- Pointnet implementation for results comparison, paper writing (21.04.2021)
- Paper writing (27.04.2021)
- Paper writing, testing on real data (05.05.2021)