Research map / Vision-language and multimodal / CLIP and zero-shot
Better Call SAL: Towards Learning to Segment Anything in Lidar
Aljosa Osep, Tim Meinhardt, Francesco Ferroni, Neehar Peri, Deva Ramanan, Laura Leal-Taixé
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| remove unlabeled points from pseudo-labeled clouds during training (ablation) | SAL · SemanticKITTI class-agnostic | PQ: 22.2 -> 59.3 | Improved |
| add FrankenFrustum augmentation (ablation) | SAL · SemanticKITTI class-agnostic | PQ: 59.3 -> 62.5 | Improved |
| mix point clouds from different scans in FrankenFrustum (ablation) | SAL · SemanticKITTI class-agnostic | PQ: 62.5 -> 62.8 | Improved |
| increase decoder queries from 100 to 300 with pseudo labels (ablation) | SAL · SemanticKITTI class-agnostic | PQ: 53.9 -> 62.8 | Improved |
| increase queries with GT labels (ablation) | SAL · SemanticKITTI stuff classes | PQ St: +0.5 | No clear effect |
| scale pseudo-labeled data (add test set to train) (ablation) | SAL · SemanticKITTI | PQ: +2.5 PQ | Improved |
| add CLIP token distillation loss (ablation) | SAL · SemanticKITTI linear probing | PQ: 20.0 -> 33.1 | Improved |
| erode SAM masks in image domain (ablation) | SAM unprojection baseline · SemanticKITTI class-agnostic | PQ: 46.0 -> 42.2 | Hurt |
| filter SAM masks by DBSCAN overlap (ablation) | SAM unprojection baseline · SemanticKITTI class-agnostic | PQ: PQ Th +27.1, PQ St -18.6, overall 46.7 | Mixed |
| replace SAM masks with DBSCAN clusters when overlap is large (ablation) | SAM unprojection baseline · SemanticKITTI class-agnostic | PQ: +3.9 (to 48.7) | Improved |
| distill lifted CLIP features into a Lidar model instead of lifting at inference | SAM+DBS+CLIP · SemanticKITTI zero-shot LPS | PQ: 27.5 -> 33.1 | Improved |
| prompt with super classes instead of default classes | SAL zero-shot · SemanticKITTI | PQ: 24.8 -> 48.5 | Improved |
Closest papers by meaning
- Zero-Shot 4D Lidar Panoptic Segmentation (CVPR 2025 · similarity 0.90)
- LESS: Label-Efficient Semantic Segmentation for LiDAR Point Clouds (ECCV 2022 · similarity 0.89)
- TerraSeg: Self-Supervised Ground Segmentation for Any LiDAR (CVPR 2026 · similarity 0.89)
- LaserMix for Semi-Supervised LiDAR Semantic Segmentation (CVPR 2023 · similarity 0.89)
- Towards Learning to Complete Anything in Lidar (ICML 2025 · similarity 0.88)
- Three Pillars Improving Vision Foundation Model Distillation for Lidar (CVPR 2024 · similarity 0.88)
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.
