[Paper] Exploring Easy Boosts for Lidar Semantic Scene Completion

Source: arXiv - 2606.03992v1

Overview

The paper Exploring Easy Boosts for Lidar Semantic Scene Completion shows that you don’t need a brand‑new network to get big gains on lidar‑based semantic scene completion (SSC). By simply enriching the raw point cloud with cheap semantic “pseudo‑labels” from existing segmentors and adding a visibility flag that tells the model what space is empty versus unknown, the authors lift the performance of several classic SSC pipelines to, and sometimes beyond, the current state‑of‑the‑art.

Key Contributions

• Semantic priors as plug‑and‑play boosters – Demonstrates that feeding off‑the‑shelf semantic predictions into any SSC model yields a large mIoU increase.
• Visibility channel – Introduces a lightweight binary mask that distinguishes observed empty voxels from truly unknown regions, giving a consistent secondary boost.
• Comprehensive oracle analysis – Quantifies how much of the improvement comes from perfect semantic priors versus other factors.
• Revival of older architectures – Shows that classic SSC networks (e.g., SSCNet, LMSCNet) become competitive again when equipped with the proposed priors, sometimes surpassing newer methods.
• Open‑source implementation – Provides a ready‑to‑use codebase (https://github.com/astra-vision/SSC-Priors) for reproducibility and rapid experimentation.

Methodology

1. Baseline SSC models – The study starts from three widely used lidar‑SSC architectures (SSCNet, LMSCNet, and a recent transformer‑based variant). No changes are made to their internal layers.
2. Semantic pseudo‑labels – An off‑the‑shelf 2‑D/3‑D segmentor (e.g., RangeNet++, Cylinder3D) runs on the raw lidar sweep, producing per‑point class probabilities. These are projected onto the voxel grid and concatenated to the original occupancy channel as extra input features.
3. Visibility channel – For each voxel the authors compute whether the lidar beam has passed through it (empty) or whether the voxel lies behind an observed surface (unknown). This binary flag is added as another input channel.
4. Training & evaluation – The augmented inputs are fed to the unchanged SSC networks and trained on the SemanticKITTI SSC benchmark. An “oracle” experiment replaces pseudo‑labels with ground‑truth semantics to isolate the upper bound of the prior’s impact.

The whole pipeline requires only a forward pass through a pre‑trained segmentor and a cheap visibility rasterization step—no redesign of the SSC backbone.

Results & Findings

• Semantic priors dominate: The oracle experiment (perfect semantics) yields an additional ~10 pp mIoU, confirming that high‑quality class information is the main driver.
• Visibility is a consistent secondary gain across all models, indicating that distinguishing empty from unknown space helps the network resolve occlusions.
• Older models close the gap: With the two cheap augmentations, LMSCNet even outperforms the latest transformer‑based SSC on the benchmark.

Practical Implications

• Fast performance upgrades – Teams can boost existing lidar‑SSC pipelines without retraining large models or investing in architecture research; just plug in a pre‑trained segmentor and the visibility mask.
• Cost‑effective deployment – The added computation (a single forward pass of a lightweight segmentor) is negligible compared to the full SSC inference, making it suitable for real‑time autonomous‑driving stacks.
• Modular design – The approach encourages a “prior‑first” mindset: treat semantic and geometric priors as interchangeable modules that can be swapped or upgraded independently.
• Benchmark re‑evaluation – Researchers should report SSC results both with and without such priors to ensure fair comparisons, as the baseline numbers can be dramatically shifted by these simple tricks.

Limitations & Future Work

• Reliance on external segmentors – The quality of the boost hinges on the accuracy of the pseudo‑labels; noisy segmentations could propagate errors into the SSC output.
• Domain shift – The study focuses on SemanticKITTI; performance gains may vary on other lidar datasets (e.g., nuScenes, Waymo) where segmentor training data differ.
• Visibility computation assumes perfect calibration – Mis‑aligned sensor models could produce incorrect empty/unknown masks.
• Future directions suggested by the authors include learning to refine the pseudo‑labels jointly with SSC, exploring richer geometric priors (e.g., surface normals), and extending the analysis to multimodal setups that fuse camera imagery.

Authors

• Tetiana Martyniuk
• Jonathan Seele
• Alexandre Boulch
• Gilles Puy
• Renaud Marlet
• Raoul de Charette

Paper Information

• arXiv ID: 2606.03992v1
• Categories: cs.CV, cs.RO
• Published: June 2, 2026
• PDF: Download PDF