Addressing class imbalance challenge in Semantic Segmentation of ALS data through performance analysis of RandLA-NET and PointNET ++

Accurate and efficient semantic labeling of point clouds is critical for utilizing the LiDAR data for varied applications. The voluminous nature of data coupled with its unstructured nature pose a challenge to the research community in achieving the desired accuracy. Computational complexity also plays a pivotal role while dealing with the large point cloud datasets. Deep learning based labeling approaches have been gaining popularity, owing to the accuracy it can deliver. However, the accuracy of deep learning architecture is heavily dependent on the availability of a large amount of training dataset. The class imbalance affects the classification accuracy of minor classes. This paper proposes to compare RandLA-Net, an encoder-decoder based deep learning network with the PointNet++, a point-based deep learning network for point cloud classification of significantly represented and under-represented classes. The proposed methodology has been tested on the benchmark DALES dataset, which has notable class imbalance of urban objects. Results show that RandLA-NET performs better, with an increase of 13mean IoU in comparison to POINT-NET++. The paper further discusses the performances of both approaches for labeling major and minor classes.