Research map / Point clouds / Segmentation and detection
BFANet: Revisiting 3D Semantic Segmentation with Boundary Feature Analysis
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 BFANet, a 3D point cloud semantic segmentation network on an OctFormer backbone that decouples semantic and boundary features with a boundary-semantic attention block, plus four error-specific evaluation metrics and a CUDA-parallel real-time boundary pseudo-label method.
Problem
Standard 3D semantic segmentation metrics (mIoU, mAcc, oAcc) hide specific error types such as false response, merging error, displacement, and region classification error, and existing boundary pseudo-label methods require preprocessing that is incompatible with data augmentation like mixup.
Approach
Decouple octree multi-level features into semantic and boundary branches with separate query/key/value projections, use boundary queries concatenated with semantic queries to attend and enrich semantic features, and compute boundary pseudo-labels online on GPU in parallel; also define four error-type metrics (FErr, MErr, RErr, DErr).
Key results
- On ScanNet200 hidden test set, BFANet reaches 36.0 mIoU vs OctFormer 32.6 (+3.4) and ranks 2nd on the official leaderboard excluding PTv3 with PPT.
- On ScanNet200 validation, BFANet reaches 37.3 mIoU, 46.7 mAcc, 84.3 oAcc, exceeding PTv3 (35.2 mIoU).
- On ScanNetv2 validation, BFANet reaches 78.0 mIoU vs PTv3 77.5; boundary pseudo-label computation takes 46.3 ms vs 179.2 ms for CBL (3.9x faster).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the boundary-semantic block to the OctFormer baseline (ablation) | BFANet / OctFormer · ScanNet200 validation | mIoU: 32.7 -> 36.4 | Helped |
| Add boundary prediction only to the OctFormer baseline (ablation) | BFANet / OctFormer · ScanNet200 validation | mIoU: 32.7 -> 33.7 | Helped |
| Add test-time augmentation to the B-S block variant (ablation) | BFANet / OctFormer · ScanNet200 validation | FErr: 30.1 -> 31.3 | Hurt |
| Add test-time augmentation to the B-S block variant (ablation) | BFANet / OctFormer · ScanNet200 validation | RErr50: 18.6 -> 18.1 | Helped |
| Add the boundary-semantic block (ablation) | BFANet / OctFormer · ScanNet200 validation | FErr: 33.7 -> 30.1 | Helped |
| Use BFANet instead of CeCo | ScanNet200 hidden test · ScanNet200 hidden test | mIoU: 34.0 -> 36.0 | Helped |
| Use BFANet instead of PTv3 | BFANet · ScanNetv2 validation | mIoU: 77.5 -> 78.0 | Helped |
| Use BFANet instead of PTv3 | BFANet · ScanNet200 validation | FErr: 33.8 -> 31.3 | Helped |
| Use BFANet instead of PTv3 | BFANet · ScanNet200 validation | MErr: 41.4 -> 35.9 | Helped |
| Use BFANet instead of PTv3 | BFANet · ScanNet200 validation | RErr50: 15.6 -> 18.1 | Hurt |
| Use BFANet instead of PTv3 | BFANet · ScanNet200 validation | DErr50: 29.4 -> 18.6 | Helped |
| Use parallel boundary pseudo-label calculation rather than CBL | PBPLC · ScanNet200 validation | mean time per scene: 179.2 -> 46.3 | Helped |
Surprising findings
- Test-time augmentation worsens FErr and MErr (31.3 vs 30.1 and 35.9 vs 34.7 without TTA), which the authors attribute to max-pooling ensembles and superpoint merging of small objects.
- Boundary prediction alone yields only small gains over baseline, while the full boundary-semantic block gives larger improvements, and the method shows little improvement on the RErr metric.
Limitations
- Improvement on region classification error (RErr) is limited, as the method focuses on boundary analysis.
- TTA degrades boundary metrics FErr and MErr.
- Evaluated only on indoor ScanNet datasets; outdoor scenarios left to future work.
Experimental setup
Datasets: ScanNet200 ScanNetv2
Benchmarks: ScanNet200 ScanNetv2
Metrics: mIoU mAcc oAcc FErr MErr RErr50 DErr50
Baselines: OctFormer PTv3 PTv2 PTv1 OA-CNN CeCo MinkUNet CBL
Builds on: OctFormer CBL HRNet PTv3
Closest papers by meaning
- SegGraph: Leveraging Graphs of SAM Segments for Few-Shot 3D Part Segmentation (NeurIPS 2025 · similarity 0.88)
- Mosaic3D: Foundation Dataset and Model for Open-Vocabulary 3D Segmentation (CVPR 2025 · similarity 0.87)
- LogoSP: Local-global Grouping of Superpoints for Unsupervised Semantic Segmentation of 3D Point Clouds (CVPR 2025 · similarity 0.87)
- 3D-AVS: LiDAR-based 3D Auto-Vocabulary Segmentation (CVPR 2025 · similarity 0.87)
- SUM Parts: Benchmarking Part-Level Semantic Segmentation of Urban Meshes (CVPR 2025 · similarity 0.87)
- UniDSeg: Unified Cross-Domain 3D Semantic Segmentation via Visual Foundation Models Prior (NeurIPS 2024 · 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.