Fine-Grained Semantic Information Preservation and Misclassification-Aware Loss for 3D Point Cloud

• Published in: IEEE robotics and automation letters Vol. 9; no. 9; pp. 7947 - 7954
• Main Authors: Zou, Yanmei, Lin, Xuefei, Yu, Hongshan, Yang, Zhengeng, Akhtar, Naveed
• Format: Journal Article
• Language: English
• Published: IEEE 01.09.2024
• Subjects: Decoding; Deep learning for visual perception; Feature extraction; Kernel; Point cloud compression; recognition; Semantics; Task analysis; Three-dimensional displays; visual learning
• ISSN: 2377-3766; 2377-3766
• DOI: 10.1109/LRA.2024.3404751

Encoder-Decoder structure is a popular choice in point cloud processing for dense multi-classification tasks, e.g., 3D semantic segmentation. Though existing techniques that follow this structure achieve high performance, they are known to suffer from fine-grained information loss, especially when the underlying networks are deep. To alleviate this, we propose a bilateral attention fusion module (BAFM) that leverages multi-resolution feature fusion to allow more effective fine-grained information flow in the network. Moreover, we also introduce a misclassification-aware loss (MAL) as a more potent alternative to the widely used cross-entropy (CE) loss for multi-classification tasks. MAL enables an explicit penalization of misclassification. Empirical experiments reveal that our method achieves state-of-the-art performance for several challenging datasets, such as 91.9% mAcc on ModelNet40, 88.4% OA and 87.0% mAcc on ScanObjectNN and 71.3% mIoU on S3DIS Area-5.