CBF-Net: An Adaptive Context Balancing and Feature Filtering Network for Point Cloud Classification
Qian Zang,Wenhui Diao,Kaiqiang Chen,Ling Liu,Menglong Yan,Xian Sun
DOI: https://doi.org/10.1109/jstars.2021.3106376
IF: 4.715
2021-01-01
IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Abstract:Point cloud classification is regarded as a critical task in remote sensing data interpretation, which is widely used in many fields. Recently, many proposed methods tend to develop an end-to-end network to directly operate on the raw point cloud, which has shown great power. However, most of these methods abstract local features by equally considering the neighboring points. The features learned may neglect to distinguish contributions of different points especially the edge points and outliers, leading to a coarse classification result especially for boundaries. Moreover, the extracted features are high redundant and intercorrelated with similar categories, posing difficulty in identifying classes sharing similar characteristics especially in complex scenes. Therefore, we propose an adaptive context balancing and feature filtering network (CBF-Net) to tackle the aforementioned problems. First, we introduce a balanced context encoding module to balance semantically the features of neighboring points, which can help the model learn more from the edge points and, therefore, contribute to a finer classification. Then, considering that the interference for similar classes probably causes confusion among them, a filtered feature aggregating module is proposed to filter the extracted features by mapping them into a cleaner subspace with a lower rank. We have conducted thorough experiments on the International Society for Photogrammetry and Remote Sensing 3-D labeling dataset. Experimental results show that our CBF-Net can obtain high accuracy and achieve state-of-the-art level in the categories of Powerline, Car, and Facade. In addition, we also conduct experiments on the RueMonge2014 dataset, which further reveals the strong ability of our model.
imaging science & photographic technology,remote sensing,engineering, electrical & electronic,geography, physical