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

Donghyeon Kwon, Taegyu Park, Suha KwakSubmission to ICLR 2026 · Rejected

In the world of autonomous driving and robotics, understanding a 3D environment is everything. LiDAR sensors provide the "eyes," but turning millions of raw points into meaningful categories—like "pedestrian," "sidewalk," or "car"—requires dense semantic segmentation. The catch? Manually labeling these 3D point clouds is a grueling, expensive process.

Semi-supervised learning (SSL) offers a way out by training models on a small amount of labeled data and a mountain of unlabeled data. However, SSL has a "confirmation bias" problem: models generate their own "pseudo-labels" for unlabeled data, and if those labels are wrong, the model effectively learns its own mistakes, leading to a downward spiral of error propagation.

A new paper, "REPL: Pseudo-label Refinement for Semi-supervised LiDAR Semantic Segmentation," introduces a clever way to break this cycle. Instead of just filtering out noisy labels, the authors propose a way to actually fix them.

TL;DR

REPL (Refinement of Pseudo-Labels) is a novel framework that uses masked reconstruction to identify and correct errors in pseudo-labels during semi-supervised training. By shifting the focus from discarding unreliable labels to actively refining them, REPL achieves state-of-the-art performance on major LiDAR benchmarks like nuScenes and SemanticKITTI.

What Problem Does This Paper Address?

The primary bottleneck in LiDAR semantic segmentation is the lack of annotated data. While SSL methods use a "Teacher" model to generate pseudo-labels for a "Student" model, these labels are often noisy.

Current state-of-the-art methods handle this noise using post-hoc strategies: they either filter out low-confidence predictions or reweight the loss so the model ignores "uncertain" points. The authors argue this is wasteful. By simply throwing away data, these models miss out on the contextual information those points could provide. The challenge REPL tackles is: Can we improve the intrinsic quality of pseudo-labels at the point of generation rather than just managing the fallout of poor labels?

Methods

REPL introduces a two-stage "Refiner" network that sits alongside the standard Teacher-Student architecture:

  1. Unreliable Voxel Identification (Detect): The system identifies "unreliable" voxels by looking for disagreements between the Teacher and Student models. If they predict different classes or if their confidence scores fall below an adaptive threshold, those voxels are marked for refinement.
  2. Masked Reconstruction (Correct): Inspired by Masked Autoencoders (MAE), the marked "unreliable" voxels are replaced with a learnable [MASK] token. The Refine Network then takes the original LiDAR data and the remaining "reliable" labels to reconstruct the missing (masked) labels. Essentially, it uses the surrounding context to "fill in the blanks" with more accurate predictions.

To make the Refiner robust, the authors use a few key training tricks:

  • Negative Learning: On unlabeled data, instead of telling the refiner what a point is, they tell it what it isn't (suppressing implausible classes).
  • LaserMix Augmentation: They mix labeled and unlabeled 3D scenes to force the refiner to handle a wider variety of errors and geometries.

Datasets

The authors evaluated REPL on two industry-standard outdoor LiDAR datasets:

  • nuScenes-lidarseg: A massive dataset with 16 semantic classes.
  • SemanticKITTI: A benchmark featuring 19 classes across diverse driving sequences.

Evaluation was performed across various "label ratios" (1%, 10%, 20%, and 50%) to see how well the model performs when human-annotated data is extremely scarce.

Key Insights and Ideas

  • Refinement is Mathematically "Easier": The authors provide a theoretical proof (Proposition 1) showing that refining a noisy label is a lower-entropy task than predicting a label from scratch. Having a "hint" (even a noisy one) reduces the search space for the model.
  • The "Improvement Condition": You don't need a perfect refiner to see gains. The authors derive a condition (Proposition 2) showing that as long as the rate of correct fixes outweighs the rate of new errors introduced, the model improves. Empirically, REPL easily clears this bar.
  • Context is King: By using masked reconstruction, the model learns the underlying structure of the world (e.g., "this red point is surrounded by sidewalk points, so it's probably sidewalk, not a car"). This is much more powerful than simple confidence filtering.

Key Results

REPL consistently outperformed existing methods, setting new state-of-the-art (SOTA) marks:

  • nuScenes: REPL achieved an average gain of +2.0 mIoU (mean Intersection over Union) over the previous best method (IT2).
  • SemanticKITTI: It achieved SOTA performance at the 1% and 50% label ratios, proving highly effective even when only 1% of the data is labeled.
  • Efficiency: Despite adding a refinement step, the computational overhead is moderate. The refiner adds about 0.25 seconds of latency and ~400MB of memory usage during inference, which the authors argue is a fair trade for a +9.1 mIoU jump over the supervised-only baseline.

Limitations and Future Work

  • Over-Correction: The qualitative analysis shows that REPL occasionally "over-corrects," changing a label that was actually correct into an incorrect one (though these cases are rare compared to the successful fixes).
  • Heuristic Detection: Currently, the "unreliable" voxels are identified using a relatively simple agreement heuristic. The authors suggest that more sophisticated error detection methods could lead to even greater performance gains.
  • Real-time Constraints: While the overhead is "moderate," for high-speed autonomous driving, every millisecond counts. Further optimizing the Refine Network for real-time deployment remains a future goal.

Summary updated

Abstract

Semi-supervised learning for LiDAR semantic segmentation often suffers from error propagation and confirmation bias caused by noisy pseudo-labels. To tackle this chronic issue, we introduce RePL, a novel framework that enhances pseudo-label quality by identifying and correcting potential errors in pseudo-labels through masked reconstruction, along with a dedicated training strategy. We also provide a theoretical analysis demonstrating the condition under which the pseudo-label refinement is beneficial, and empirically confirm that the condition is mild and clearly met by RePL. Extensive evaluations on the nuScenes-lidarseg and SemanticKITTI datasets show that RePL improves pseudo-label quality a lot and, as a result, achieves the state of the art in LiDAR semantic segmentation.
BibTeX
@misc{kwon2026repl,
  author = {Donghyeon Kwon and Taegyu Park and Suha Kwak},
  title = {{RePL: Pseudo-label Refinement for Semi-supervised LiDAR Semantic Segmentation}},
  year = {2026},
  eprint = {2604.06825},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2604.06825},
}
Figure