Can we Trust Unreliable Voxels? Exploring 3D Semantic Occupancy Prediction under Label Noise

TL;DR

This work establishes OccNL, the first benchmark dedicated to 3D occupancy under occupancy-asymmetric and dynamic trailing noise, and proposes DPR-Occ, a principled label noise-robust framework that constructs reliable supervision through dual-source partial label reasoning.

Abstract

3D semantic occupancy prediction is a cornerstone of robotic perception, yet real-world voxel annotations are inherently corrupted by structural artifacts and dynamic trailing effects. This raises a critical but underexplored question: can autonomous systems safely rely on such unreliable occupancy supervision? To systematically investigate this issue, we establish OccNL, the first benchmark dedicated to 3D occupancy under occupancy-asymmetric and dynamic trailing noise. Our analysis reveals a fundamental domain gap: state-of-the-art 2D label noise learning strategies collapse catastrophically in sparse 3D voxel spaces, exposing a critical vulnerability in existing paradigms. To address this challenge, we propose DPR-Occ, a principled label noise-robust framework that constructs reliable supervision through dual-source partial label reasoning. By synergizing temporal model memory with representation-level structural affinity, DPR-Occ dynamically expands and prunes candidate label sets to preserve true semantics while suppressing noise propagation. Extensive experiments on SemanticKITTI demonstrate that DPR-Occ prevents geometric and semantic collapse under extreme corruption. Notably, even at 90% label noise, our method achieves significant performance gains (up to 2.57% mIoU and 13.91% IoU) over existing label noise learning baselines adapted to the 3D occupancy prediction task. By bridging label noise learning and 3D perception, OccNL and DPR-Occ provide a reliable foundation for safety-critical robotic perception in dynamic environments. The benchmark and source code will be made publicly available at https://github.com/mylwx/OccNL.

Figures (3)

• Figure 1: Semantic distribution evolution under voxel-level category-flipping noise. Due to orders-of-magnitude differences in voxel counts, we employ a logarithmic scale for visualization (empty voxels omitted as they remain constant at $10^{-3}\eta$ flip rate). Increased noise drives the distribution toward uniformity: dominant classes (e.g., vegetation, road) are suppressed, while rare classes (e.g., motorcyclist, person) are artificially inflated, culminating in nearly identical voxel frequencies at $90\%$ noise.
• Figure 2: The overall framework of our proposed DPR-Occ. Warm-up Stage: The model captures clean patterns via standard training on noisy labels $\tilde{Y}$. Robust Stage: Guided by dynamic-$K$ scheduling, we construct dual-source partial label sets by fusing Top-$K$ predictions from the EMA teacher and feature-prototype similarities. The network is then optimized using Partial Label Learning (PLL) and Negative Learning (NL), with EMA-guided Self-Not-True Distillation (SNTD) further suppressing noise for robust learning.
• Figure 3: Qualitative results under $90\%$ asymmetric noise on OccNL benchmark. Compared to collapsing baselines, DPR-Occ preserves structural integrity and reliable semantics. The final column shows a failure case (yellow box), where DPR-Occ still reconstructs basic road and vegetation geometry despite semantic misclassification.