Small but mighty: Enhancing 3D point clouds semantic segmentation with U-Next framework

Ziyin Zeng, Qingyong Hu, Zhong Xie, Bijun Li, Jian Zhou, Yongyang Xu

Abstract: Highlights•We identify U-Net L1<math><msup is="true"><mrow is="true"><mi is="true">L</mi></mrow><mrow is="true"><mn is="true">1</mn></mrow></msup></math> sub-network as a suitable component for point cloud segmentation.•The U-Next designed by stacking U-Net L1<math><msup is="true"><mrow is="true"><mi is="true">L</mi></mrow><mrow is="true"><mn is="true">1</mn></mrow></msup></math> to learn features from different scales.•A multi-level deep supervision mechanism introduced for smoothing gradient propagation.•Our method achieves excellent performance on large-scale benchmarks.•Our method can be integrated into existing segmentation network..