Backward Attentive Fusing Network With Local Aggregation Classifier for 3D Point Cloud Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Analysis

Background:

• 3D point cloud semantic segmentation is crucial for understanding complex environments.
• Existing methods often struggle with semantic gaps and context-awareness.
• Improving segmentation accuracy and smoothness remains a key challenge.

Purpose of the Study:

• To introduce a novel network, BAF-LAC, for enhanced 3D point cloud semantic segmentation.
• To address the semantic gap between encoder and decoder features.
• To improve context-awareness within the classification process.

Main Methods:

• Developed a Backward Attentive Fusing Encoder-Decoder (BAF-ED) to fuse multi-layer features.
• Utilized an attention mechanism to modulate low-level features with high-level ones.
• Introduced a Local Aggregation Classifier (LAC) for adaptive feature enhancement and context retention.

Main Results:

• BAF-LAC effectively narrows the semantic gap between encoder and decoder.
• The Local Aggregation Classifier preserves context consistency during classification.
• The method demonstrates improved discriminative feature extraction and smoother segmentation predictions.

Conclusions:

• The proposed BAF-LAC method significantly advances 3D point cloud semantic segmentation.
• Experimental results on Semantic3D, SemanticKITTI, and S3DIS datasets show competitive performance.
• BAF-LAC offers a promising approach for state-of-the-art semantic segmentation.