A study on local feature descriptors for point clouds
I proposed and tested some local feature descriptors for point clouds. My main goal was to handcraft features that performed well (in terms of descriptiveness) on high clutter and noisy scenarios, such as pattern recognition on Kinect point clouds. I found out that even the state-of-the-art descriptors which have excelent performance on object retrieval tasks performed poorly, and the features I proposed could at best perform close to these poor results but being faster. As a subproduct of this project, I created a small platform based on Point Cloud Library to benchmark the descriptiveness of various feature descriptors on the “classic” datasets, based on the work of Guo et al.