Semantic Segmentation for Full-Waveform LiDAR Data Using Local and Hierarchical Global Feature Extraction
Published:
- We introduce a novel waveform-aware convolutional method that directly applies convolutions on irregular full-waveform LiDAR data to extract waveform features.
- We develop an encoder-decoder-based network with a global module including downsampling and upsampling blocks with a skip connection and waveform-aware convolutional operations and a local module with only waveform-aware convolutional operations.
- We eliminate the requirement of expensive calculations of handcrafted features and achieve superior performance on a benchmark dataset without any conversion of full-waveform LiDAR data to 2D images or voxels