A TWO-STEP FEATURE EXTRACTION ALGORITHM: APPLICATION TO DEEP LEARNING FOR POINT CLOUD CLASSIFICATION

Abstract. Most deep learning (DL) methods that are not end-to-end use several multi-scale and multi-type hand-crafted features make the network challenging, more computationally intensive vulnerable to overfitting. Furthermore, reliance on empirically-based feature dimensionality reduction may lead misclassification. In contrast, efficient management can reduce storage computational complexities, builds better classifiers, improves overall performance. Principal Component Analysis (PCA) is a well-known dimension technique has been used for extraction. This paper presents two-step PCA based extraction algorithm employs variant of feature-based PointNet (Qi et al., 2017a) point cloud classification. extends framework large-scale aerial LiDAR data, contributes by (i) developing new algorithm, (ii) exploring impact in extraction, (iii) introducing non-end-to-end per classification clouds. demonstrated laser scanning (ALS) The successfully reduces space without sacrificing performance, as benchmarked against original algorithm. When tested Vaihingen data set, proposed achieves an Overall Accuracy (OA) 74.64% using 9 input vectors 14 shape features, whereas with same only 5PCs (principal components built features) it actually higher OA 75.36% which demonstrates effect reduction.