Research map / Point clouds / Point cloud classification
Fully Convolutional Geometric Features
Christopher Choy, Jaesik Park, Vladlen Koltun
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 |
|---|---|---|---|
| use hardest-negative mining in contrastive loss instead of random contrastive (ablation) | FCGF contrastive loss · 3DMatch 5cm | feature-match recall: 0.7309 -> 0.9344 | Improved |
| use hardest-contrastive loss instead of normalized contrastive loss (ablation) | FCGF · 3DMatch 5cm | feature-match recall: 0.8493 -> 0.9344 | Improved |
| use hardest-contrastive instead of best hardest-triplet mixture (ablation) | FCGF · 3DMatch 5cm | feature-match recall: 0.9087 -> 0.9344 | Improved |
| add hardest negatives to random triplets (128 HN with 2048 RT) (ablation) | triplet loss · 3DMatch 5cm | feature-match recall: 0.7903 -> 0.9087 | Improved |
| increase number of hardest negatives to 1024 with 512 random triplets (ablation) | hardest-triplet loss · 3DMatch 5cm | feature-match recall: COLLAPSE | Hurt |
| normalize features in contrastive loss (ablation) | contrastive loss · 3DMatch 5cm | feature-match recall: 0.7309 -> 0.8493 | Improved |
| normalize features in triplet loss (ablation) | triplet loss · 3DMatch 5cm | feature-match recall: 0.7903 -> 0.6935 | Hurt |
| increase feature dimension from 16 to 64 (ablation) | FCGF hardest-contrastive · 3DMatch 5cm | feature-match recall: 0.9011 -> 0.9343 | Improved |
| increase feature dimension from 32 to 48 (ablation) | FCGF hardest-contrastive · 3DMatch 5cm | feature-match recall: 0.9242 -> 0.9235 | No clear effect |
| use 64-dim features at 2.5cm voxel (ablation) | FCGF hardest-contrastive · 3DMatch 2.5cm | feature-match recall: OOM | Hurt |
| raise negative margin from 0.5 to 2 at positive margin 0.1 (ablation) | FCGF hardest-contrastive · 3DMatch 5cm | feature-match recall: 0.8832 -> 0.9344 | Improved |
| increase positive margin from 0.1 to 0.4 at negative margin 1 (ablation) | FCGF hardest-contrastive · 3DMatch 5cm | feature-match recall: 0.9249 -> 0.9013 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- PPFNet — PPFNet: Global Context Aware Local Features for Robust 3D Point Matching (CVPR 2018) · better on 2 of 2 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
- Learning Compact Geometric Features (ICCV 2017 · similarity 0.90)
- PointCNN++: Performant Convolution on Native Points (CVPR 2026 · similarity 0.88)
- Modeling Local Geometric Structure of 3D Point Clouds Using Geo-CNN (CVPR 2019 · similarity 0.88)
- Fully-Convolutional Point Networks for Large-Scale Point Clouds (ECCV 2018 · similarity 0.88)
- Hierarchical Metric Learning and Matching for 2D and 3D Geometric Correspondences (ECCV 2018 · similarity 0.88)
- GAPrompt: Geometry-Aware Point Cloud Prompt for 3D Vision Model (ICML 2025 · similarity 0.87)
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.
