Parameters for training on an S3DIS-like office building dataset #269
Description
Dear Mr. Thomas,
thank you very much for your excellent work on KPConv and for providing the PyTorch implementation. For my master’s thesis, I plan to apply your framework to a custom dataset.
As training data, I use a self-acquired point cloud of an office building with typical office rooms. For my research, I restrict the semantic segmentation to four classes: wall, ceiling, floor, and a fourth aggregated class that contains all remaining objects (e.g., windows, tables, chairs, and other furnishings commonly found in office environments).
The dataset was acquired using a mobile laser scanning system (NavVis VLX) and comprises approximately 300 million points in total. The spatial extent of the dataset can be roughly estimated from the scale bar shown in the attached figure.
I have already successfully set up your code and am able to reproduce results on the S3DIS dataset that are comparable to those reported in your paper. The training is conducted using the PyTorch version of KPConv.
If you require any additional information about my dataset or the selected parameters, I would be happy to provide it.
Kind regards
