Research map / Point clouds / Segmentation and detection
Geodesic-Former: A Geodesic-Guided Few-Shot 3D Point Cloud Instance Segmenter
Tuan Ngo, Khoi Nguyen
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 |
|---|---|---|---|
| replace cosine similarity with a similarity network (ablation) | DyCo3D baseline · ScanNetV2 val (1-shot) | mAP: 6.2 -> 6.6 | Improved |
| replace clustering with transformer decoder (ablation) | DyCo3D baseline · ScanNetV2 val (1-shot) | mAP: 6.2 -> 6.7 | Improved |
| add geodesic distance embedding to dynamic convolution (ablation) | DyCo3D baseline · ScanNetV2 val (1-shot) | mAP: 6.2 -> 7.8 | Improved |
| add geodesic embedding to transformer decoder variant (ablation) | Transformer decoder · ScanNetV2 val (1-shot) | mAP: 6.7 -> 9.4 | Improved |
| add similarity network to transformer decoder + geodesic embedding (ablation) | Geodesic-Former · ScanNetV2 val (1-shot) | mAP: 9.4 -> 10.6 | Improved |
| add similarity network to transformer decoder + geodesic embedding (ablation) | Geodesic-Former · ScanNetV2 val (1-shot) | AP50: 17.1 -> 19.8 | Improved |
| use classification head instead of similarity network (ablation) | Geodesic-Former · ScanNetV2 val (1-shot) | mAP: 10.6 -> 4.5 | Hurt |
| use per-point classification baseline instead of DyCo3D clustering (ablation) | DyCo3D baseline · ScanNetV2 val (1-shot) | mAP: 6.2 -> 4.9 | Hurt |
| use 2 dynamic conv layers instead of 1 (ablation) | Geodesic-Former · ScanNetV2 test set | mAP: 3.4 -> 10.6 | Improved |
| use 4 dynamic conv layers instead of 2 (ablation) | Geodesic-Former · ScanNetV2 test set | mAP: 10.6 -> 6.4 | Hurt |
| ball query kappa=64 instead of 16 (tau 0.05 m) (ablation) | Geodesic-Former · ScanNetV2 val | mAP: 9.2 -> 10.6 | Improved |
| increase kappa from 64 to 128 (tau 0.05 m) (ablation) | Geodesic-Former · ScanNetV2 val | mAP: 10.6 -> 10.6 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- DyCo3D — DyCo3D: Robust Instance Segmentation of 3D Point Clouds Through Dynamic Convolution (CVPR 2021) · better on 8 of 8 shared results
- PointGroup — PointGroup: Dual-Set Point Grouping for 3D Instance Segmentation (CVPR 2020) · better on 8 of 8 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DyCo3D: Robust Instance Segmentation of 3D Point Clouds Through Dynamic Convolution (CVPR 2021 · similarity 0.91)
- OneFormer3D: One Transformer for Unified Point Cloud Segmentation (CVPR 2024 · similarity 0.90)
- FreePoint: Unsupervised Point Cloud Instance Segmentation (CVPR 2024 · similarity 0.90)
- PointInst3D: Segmenting 3D Instances by Points (ECCV 2022 · similarity 0.90)
- Divide and Conquer: 3D Point Cloud Instance Segmentation With Point-Wise Binarization (ICCV 2023 · similarity 0.89)
- Dynamic Convolution for 3D Point Cloud Instance Segmentation (TPAMI 2022 · similarity 0.89)
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.
