MEHPool: Multispectral airborne LiDAR point cloud classification with maximum entropy hierarchical pooling
Deep learning methods for point cloud classification often struggle with two things: (1) how to select representative neighbor points for local feature extraction, and (2) how to pool those features without losing too much information.
Common pooling methods (max, average, etc.) are simple but tend to discard fine-grained local details. And most neighbor selection strategies either take too few points (missing context) or too many (bringing in noise).
So we proposed MEHPool, a plug-and-play module that does two things:
(1) Homogeneous Neighbor Selection (HNS): uses the maximum entropy principle to adaptively select neighbor points that are actually similar to the core point, which reduces interference from other classes.
(2) Graph Pooling (GP): then pools those selected points into smaller, cleaner graphs while preserving the geometric structure. The idea is to keep the most informative vertices and drop the rest.
while preserving the geometric structure. The idea is to keep the most informative vertices and drop the rest.
If you find our work useful for your research, please cite:
@article{jiang2024multispectral,
title={Multispectral airborne LiDAR point cloud classification with maximum entropy hierarchical pooling},
author={Jiang, Ge and Lichti, Derek D and Yin, Tiangang and Yan, Wai Yeung},
journal={IEEE Geoscience and Remote Sensing Letters},
volume={22},
pages={1--5},
year={2024},
publisher={IEEE}
}