RePL: Pseudo-label Refinement for Semi-supervised LiDAR Semantic Segmentation

The paper introduces RePL, a framework that refines pseudo-labels in semi-supervised LiDAR semantic segmentation through masked reconstruction and a dedicated training strategy. It resides in the Pseudo-Label Quality Enhancement leaf, which contains only two papers including the original work. This leaf sits under Pseudo-Label Generation and Refinement, a branch with four total papers. The sparse population suggests this specific focus on explicit pseudo-label correction mechanisms is relatively underexplored compared to broader consistency-based or weakly supervised directions, which collectively contain over thirty papers across multiple leaves.

The taxonomy reveals that neighboring research directions pursue complementary strategies. Consistency-Based Semi-Supervised Learning (nine papers across three leaves) enforces prediction invariance under spatial, temporal, or contrastive perturbations, while Cross-Modal and Multi-Modal Learning (ten papers) leverages image data to guide LiDAR segmentation. Weakly Supervised Learning with Reduced Annotations (eighteen papers across four leaves) explores alternative supervision signals like scribbles or bounding boxes. The scope note for Pseudo-Label Quality Enhancement explicitly excludes methods that generate pseudo-labels without refinement, positioning RePL in a narrower niche focused on error correction rather than initial label production or consistency regularization.

Among fourteen candidates examined, none clearly refute any of the three contributions. The core RePL framework examined four candidates with zero refutable overlaps, the theoretical analysis examined ten candidates with zero refutable overlaps, and the training strategy examined zero candidates. This limited search scope—covering top-K semantic matches and citation expansion—suggests that within the examined literature, no prior work explicitly combines masked reconstruction for pseudo-label refinement with theoretical guarantees in this domain. However, the small candidate pool and sparse taxonomy leaf indicate that the search may not have captured all relevant pseudo-labeling or self-training methods from adjacent computer vision subfields.

Based on the available signals, RePL appears to occupy a relatively sparse research direction within semi-supervised LiDAR segmentation, with limited direct prior work on masked reconstruction for pseudo-label correction. The analysis covers fourteen candidates from top-K semantic search, not an exhaustive survey of all pseudo-labeling or self-training literature. The theoretical contribution and training strategy show no refutable overlaps in the examined set, though the narrow search scope and sparse taxonomy leaf suggest caution in generalizing these findings beyond the specific LiDAR segmentation context.