Research map / Detection and segmentation / Open-set and few-shot detection
No Thing, Nothing: Highlighting Safety-Critical Classes for Robust LiDAR Semantic Segmentation in Adverse Weather
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 NTN, which adds superclass feature binding and beam-wise feature distillation on top of SJ+LPD to improve LiDAR semantic segmentation of "things" classes under adverse weather in a domain generalization setting.
Problem
Domain generalization LiDAR semantic segmentation under adverse weather performs much worse on safety-critical "things" classes than on "stuff" classes, with things often misclassified as stuff.
Approach
Binds each point feature to a superclass prototype (grouping visually similar classes) to prevent things-to-stuff confusion, and distills clean-weather features into corrupted features averaged per LiDAR beam for things classes only, to compensate for local point loss.
Key results
- NTN reaches 38.9 mIoU on SemanticKITTI-to-SemanticSTF, +2.6 over SJ+LPD (36.3), with things mIoU up 4.8.
- On SemanticPOSS-to-SemanticSTF, NTN reaches 46.2 mIoU, +7.9 over SJ+LPD (38.3).
- Ablation shows superclass prototypes beat class-wise prototypes by +3.3 things mIoU, and beam-wise distillation on things only beats point-wise distillation (34.4 vs 33.1 things).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Feature Binding and Beam-wise Feature Distillation to SJ+LPD | SemanticKITTI-to-SemanticSTF · SemanticKITTI-to-SemanticSTF | mIoU: 36.3 -> 38.9 (+2.6) | Helped |
| Add Feature Binding and Beam-wise Feature Distillation to SJ+LPD | SemanticPOSS-to-SemanticSTF · SemanticPOSS-to-SemanticSTF | mIoU: 38.3 -> 46.2 (+7.9) | Helped |
| Add superclass prototypes instead of using class-wise prototypes (ablation) | Feature Binding on SemanticKITTI-to-SemanticSTF · SemanticKITTI-to-SemanticSTF | things mIoU: 31.6 -> 34.9 (↑3.3) | Helped |
| Use class-wise prototypes instead of superclass prototypes (ablation) | Feature Binding on SemanticKITTI-to-SemanticSTF · SemanticKITTI-to-SemanticSTF | mIoU: 38.1 -> 36.1 | Hurt |
| Use two coarse things/stuff prototypes instead of superclass prototypes (ablation) | Feature Binding on SemanticKITTI-to-SemanticSTF · SemanticKITTI-to-SemanticSTF | mIoU: 38.1 -> 37.9 | Hurt |
| Apply beam-wise feature distillation to things classes rather than all classes (ablation) | Feature Distillation on SemanticKITTI-to-SemanticSTF · SemanticKITTI-to-SemanticSTF | mIoU: 37.9 -> 38.4 | Helped |
| Use point-wise instead of beam-wise feature distillation on things (ablation) | Feature Distillation on SemanticKITTI-to-SemanticSTF · SemanticKITTI-to-SemanticSTF | mIoU: 38.4 -> 36.7 | Hurt |
| Combine superclass prototypes with beam-wise distillation on things (ablation) | NTN on SemanticKITTI-to-SemanticSTF · SemanticKITTI-to-SemanticSTF | things mIoU: 31.4 -> 36.2 (↑4.8) | Helped |
| Combine superclass prototypes with beam-wise distillation on things (ablation) | NTN on SemanticKITTI-to-SemanticSTF · SemanticKITTI-to-SemanticSTF | mIoU: 36.3 -> 38.9 (↑2.6) | Helped |
Surprising findings
- Coarse things/stuff two-prototype binding improves things classes but reduces stuff performance enough to lower overall mIoU versus superclass prototypes.
- Applying beam-wise distillation to all classes performed worse than restricting it to things classes, attributed to class imbalance from stuff classes.
Limitations
- Gains on things classes are uneven; some things classes (e.g., bicycle, traffic-sign shifts) show decreases or small changes.
- Evaluated only with MinkowskiNet-18 backbone and SJ+LPD-based pipeline; architecture independence is claimed but not tested.
- Relies on off-the-shelf beam separation algorithms and manually defined superclass groupings.
Experimental setup
Datasets: SemanticKITTI SemanticPOSS SemanticSTF
Benchmarks: SemanticKITTI-to-SemanticSTF SemanticPOSS-to-SemanticSTF
Metrics: mIoU IoU per class
Baselines: Source-only Dropout Perturbation PolarMix MMD PCL PointDR DGLSS UniMix DGUIL SJ+LPD
Builds on: SJ+LPD MinkowskiNet
Closest papers by meaning
- UniMix: Towards Domain Adaptive and Generalizable LiDAR Semantic Segmentation in Adverse Weather (CVPR 2024 · similarity 0.92)
- Towards Explicit Geometry-Reflectance Collaboration for Generalized LiDAR Segmentation in Adverse Weather (CVPR 2025 · similarity 0.89)
- Parsing All Adverse Scenes: Severity-Aware Semantic Segmentation with Mask-Enhanced Cross-Domain Consistency (AAAI 2024 · similarity 0.88)
- LiON: Learning Point-Wise Abstaining Penalty for LiDAR Outlier DetectioN Using Diverse Synthetic Data (AAAI 2025 · similarity 0.87)
- Semantic Segmentation in Multiple Adverse Weather Conditions with Domain Knowledge Retention (AAAI 2024 · similarity 0.87)
- An Empirical Study of the Generalization Ability of Lidar 3D Object Detectors to Unseen Domains (CVPR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.