Auto-labeling for single-photon LiDAR semantic understanding under varying acquisition conditions.

Abstract

Auto-labeling has recently emerged as a promising direction for enlarging labeled datasets under explicit error control by iteratively expanding a small human-labeled set with selectively trusted model predictions. A key challenge, however, is that prediction accuracy and confidence calibration can vary with sample-wise measurement quality, so that the same confidence score may correspond to different error rates across measurement quality ranges. As a result, applying a single global confidence threshold can introduce structured selection bias, causing low-quality measurements to accumulate disproportionately many labeling errors and degrading downstream learning. To address this issue, we propose Kernel-Weighted Auto-Labeling (KWAL), a condition-aware auto-labeling framework that leverages per-sample acquisition statistics when calibrating confidence and selecting pseudo-labels. KWAL trains a lightweight calibration model on classifier outputs together with acquisition information, and uses a kernel-weighted objective to encourage balanced calibration across acquisition ranges. It further promotes robustness in challenging conditions through an anchor-wise conditional value-at-risk loss and stabilizes decisions via historical prediction aggregation across iterations. We instantiate KWAL on single-photon LiDAR, where photon statistics provide informative acquisition indicators. Experiments on depth-image classification (synthetic and real) and point cloud semantic segmentation show that KWAL achieves high-coverage auto-labeling with substantially more consistent accuracy across photon-count (measurement-quality) ranges, reducing the accuracy gap between low- and high-quality conditions by more than half compared with existing auto-labeling baselines.