# Dual-Branch Point Cloud Semantic Segmentation: An EMA-Based Teacher–Student Collaborative Learning Framework

## TL;DR

This paper introduces a new framework for point cloud segmentation that improves performance with very few labeled examples.

## Contribution

The novel dual-branch consistency learning framework uses EMA and combines prediction and feature-level regularization for better semi-supervised learning.

## Key findings

- DBCL achieves 68.56% mIoU on S3DIS with only 0.1% labels.
- The method outperforms existing semi-supervised approaches and matches some fully supervised baselines.
- The framework uses JS divergence and contrastive learning for prediction and feature alignment.

## Abstract

Point cloud semantic segmentation remains challenging under extremely low annotation budgets due to inefficient utilization of sparse labels and sensitivity to data augmentation noise. To address this, we propose a dual-branch consistency learning (DBCL) framework featuring an EMA teacher for semi-supervised point cloud segmentation. Our core innovation lies in a unified consistency regularization scheme that enforces prediction-level alignment via JS divergence and feature-level contrastive learning, while a geometry-aware Laplacian smoothing term preserves local structural consistency. Extensive experiments demonstrate that DBCL achieves 68.56% mIoU on S3DIS with only 0.1% labels, outperforming existing semi-supervised methods and even matching some fully supervised baselines.