I'm researching surface reconstruction from point clouds using machine learning. I'd like to have some possible optimizations investigated for the reconstruction. On top, there is a very interesting idea for automatically tuning parameters...
- (*) Optimize the speed and quality of the reconstruction using implicit surfaces. Some of these:
- Scene-based training data with multiple objects per training sample.
- Add training data far from the surface.
- Local and global information:
- Re-enable uniform random sampling for the point cloud sub-samples.
- More intelligent sub-sampling. Extra neural networks?
- Multi-scale approach for local and global information.
- Better sign propagation of the sparse distance field.
- Find better ways to add sample distances to the grid, e.g. with Gaussian smoothing.
- Evaluate query points multiple times and take the average.
- Find better query points, e.g. with iterative refining.
- Combine input point clouds with images
- Optimize parameters.
- Data augmentation: translate point clouds further over the borders of the volume and clip points.
- (* for DA) (Semi)-Automatic Parameter-Space Exploration: The reconstruction is done multiple times with random parameters. Then, it optimizes the parameters with e.g. Gaussian Processes for better results.
- Post-processing, e.g. mesh smoothing to get rid of artifacts.
- Make the reconstruction available via webserver.
The choice of tasks depends on type of thesis or project and your preferences. Tasks with * are mandatory.
- Knowledge of English language (source code comments and final report should be in English)
- Basic knowledge of geometry for computer graphics (e.g. surface definition with vertices and faces)
- Basic knowledge of Python
- Basic knowledge of Deep Learning (Pytorch)
- Experience with C++
- Knowledge of modeling and geometry processing
The source code is currently pure Python for both Windows and Linux. C++ might be necessary for performance optimization.