Research map / Point clouds / Segmentation and detection

Towards Explicit Geometry-Reflectance Collaboration for Generalized LiDAR Segmentation in Adverse Weather

CVPR 2025 · Point clouds › Segmentation and detection · approach: LiDAR segmentation for driving · 5 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes GRC/GRCNet, a dual-branch LiDAR segmentation network that separately encodes geometry (voxel sparse conv) and reflectance (range-view 2D conv) and fuses them via variational-style information constraint and local/global fusion, for domain generalization to adverse weather without augmentation or simulation.

Problem

Generalized LiDAR semantic segmentation degrades under adverse weather, and prior work largely ignores that weather causes heterogeneous domain shifts in point-cloud geometry versus reflectance intensity.

Approach

Explicitly separate geometry and reflectance feature encoding into two branches (voxel 3D sparse conv and range-view 2D conv), then fuse them with a Complementarity-aware Information Constraint (variational KL-based) and a robust multi-level fusion (dynamic local weighting by feature std plus learnable-query two-stage cross-attention global fusion), instead of relying on weather simulation or data

Key results

• On SemanticKITTI→SemanticSTF, GRCNet reaches 42.5 mIoU versus 24.4 for the MinkNet baseline (+18.1) and 39.5 for RDA.
• On SynLiDAR→SemanticSTF, the method reaches 23.5 mIoU, slightly above UniMix (23.4) without weather simulation or EMA.
• With SPVCNN and MinkNet18/32 backbones, the method improves SemanticKITTI-C mIoU to 54.2 and 60.3, respectively.

Measured effects reported in the paper

Surprising findings

• Simply removing reflectance intensity from the input improves MinkNet's SemanticSTF mIoU from 24.4 to 37.3, indicating reflectance hurts generalization under adverse weather.
• Naively adding a reflectance encoder's features to the geometry branch lowers overall mIoU by 0.8 (37.3 to 36.5).

Limitations

• Adding reflectance naively degrades performance, so the benefit depends on the specific fusion design.
• Cross-attention between the two full feature maps runs out of memory on an RTX 4090, so the proposed two-stage design is needed for efficiency.
• Adding RMFC to shallower encoder blocks reduces target-domain performance, likely due to noise in shallow features.

Experimental setup

Datasets: SemanticKITTI SynLiDAR SemanticSTF SemanticKITTI-C

Benchmarks: SemanticSTF SemanticKITTI-C

Metrics: mIoU

Baselines: MinkNet PolarMix MMD PCL PointDR UniMix RDA SPVCNN

Builds on: MinkowskiNet RDA PointDR variational information bottleneck

Closest papers by meaning

• BeyondSparse: Facilitating Mamba to Enhance Cross-Domain 3D Semantic Segmentation in Adverse Weather (AAAI 2026 · similarity 0.89)
• No Thing, Nothing: Highlighting Safety-Critical Classes for Robust LiDAR Semantic Segmentation in Adverse Weather (CVPR 2025 · similarity 0.89)
• UniMix: Towards Domain Adaptive and Generalizable LiDAR Semantic Segmentation in Adverse Weather (CVPR 2024 · similarity 0.88)
• STGC-NeRF: Spatial-Temporal Geometric Consistency for LiDAR Neural Radiance Fields in Dynamic Scenes (AAAI 2025 · similarity 0.86)
• Efficient LiDAR Reflectance Compression via Scanning Serialization (ICML 2025 · similarity 0.85)
• PC-BEV: An Efficient Polar-Cartesian BEV Fusion Framework for LiDAR Semantic Segmentation (AAAI 2025 · similarity 0.85)