Research map / Label-efficient and robust learning / Domain adaptation and generalization

UniMix: Towards Domain Adaptive and Generalizable LiDAR Semantic Segmentation in Adverse Weather

CVPR 2024 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Domain-generalized segmentation · 32 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 UniMix, a two-stage teacher-student method that uses physically based weather simulation (a Bridge Domain) and a Universal Mixing operator (spatial, intensity, semantic) to adapt and generalize LiDAR semantic segmentation from clear to adverse weather.

Problem

LiDAR semantic segmentation models trained on clear-weather data degrade badly in adverse weather and unseen scenes, with domain shift from both scene layout and weather-induced intensity/spatial changes.

Approach

Construct a Bridge Domain by physically simulating fog, rain, and snow on source scans, then mix point clouds across domains via a Universal Mixing operator (spatial, intensity, and semantic masks) in a two-stage teacher-student pipeline (source-to-bridge, bridge-to-target), which serves both UDA and DG.

Key results

• UDA on SemanticKITTI→SemanticSTF reaches 37.2 mIoU vs 24.4 source-only and 28.4 for CoSMix (gain of 12.8 over source-only).
• DG on SemanticKITTI→SemanticSTF reaches 31.4 mIoU, a gain of 7.0 over source-only and above PointDR's 28.6 (gain of 4.2).
• On SynLiDAR→SemanticKITTI, UniMix achieves 33.7 mIoU, a 13.3 gain over source-only.

Measured effects reported in the paper

Surprising findings

• Light fog as the Bridge Domain gave the smallest gain, and the mixed-fog setting gave the largest, which the authors attribute to diversity of perturbation patterns.
• Semantic mixing gave the largest whole-set gain in the mixing ablation but performed worse on the rain subset than spatial or intensity mixing.

Limitations

• Adverse weather is limited to four simulated conditions (dense fog, light fog, rain, snow) and the paper notes that light-fog simulation gives small gains.
• Synthetic-to-real adaptation is harder than real-to-real adaptation, and results on several classes (e.g., bicycle, motorcycle) remain low.
• Relies on a teacher-student pseudo-labeling pipeline whose quality on the target domain is not separately analyzed.

Experimental setup

Datasets: SemanticKITTI SynLiDAR SemanticSTF STF

Benchmarks: SemanticKITTI→SemanticSTF SynLiDAR→SemanticSTF SynLiDAR→SemanticKITTI

Metrics: mIoU per-class IoU

Baselines: ADDA Ent-Min Self-training CoSMix Dropout PolarMix MMD PCL PointDR LaserMix

Builds on: CoSMix PolarMix LaserMix PointDR MinkowskiNet

Closest papers by meaning

• No Thing, Nothing: Highlighting Safety-Critical Classes for Robust LiDAR Semantic Segmentation in Adverse Weather (CVPR 2025 · similarity 0.92)
• Parsing All Adverse Scenes: Severity-Aware Semantic Segmentation with Mask-Enhanced Cross-Domain Consistency (AAAI 2024 · similarity 0.89)
• Semantic Segmentation in Multiple Adverse Weather Conditions with Domain Knowledge Retention (AAAI 2024 · similarity 0.89)
• Towards Explicit Geometry-Reflectance Collaboration for Generalized LiDAR Segmentation in Adverse Weather (CVPR 2025 · similarity 0.88)
• An Empirical Study of the Generalization Ability of Lidar 3D Object Detectors to Unseen Domains (CVPR 2024 · similarity 0.87)
• LiT: Unifying LiDAR "Languages" with LiDAR Translator (NeurIPS 2024 · similarity 0.87)