Research map / Detection and segmentation / Weakly supervised detection and segmentation
BSNet: Box-Supervised Simulation-assisted Mean Teacher for 3D Instance Segmentation
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 BSNet, a box-supervised 3D instance segmentation method whose pseudo-labeler SAFormer combines Mean Teacher with simulated overlapping samples and a Local-Global Aware Attention decoder to label overlapping bbox regions.
Problem
Bounding-box-supervised 3D instance segmentation suffers from indeterminate instance labels for points in overlapping boxes, so pseudo-labels for those regions are unreliable.
Approach
Builds simulated overlapping samples from non-overlapping boxes (matched to real overlap statistics, with gravity and collision constraints) to pretrain a pseudo-labeler, then fine-tunes it inside a Mean Teacher framework with a Local-Global Aware Attention decoder that models local instance structure and global context.
Key results
- On ScanNetV2 validation, Ours + SPFormer reaches 53.3 mAP and 72.7 AP@50, versus 51.1 mAP for GaPro + SPFormer.
- On S3DIS Area 5, Ours + ISBNet reaches 53.0 mAP and 64.3 AP@50, versus 50.5 mAP for GaPro + ISBNet.
- Pseudo-label overlap-area mAcc improves from 38.1 (Gapro) to 59.6 with the full SAFormer pipeline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BSNet pseudo-labels with SPFormer instead of GaPro pseudo-labels | SPFormer · ScanNetV2 validation | mAP: 51.1 -> 53.3 | Helped |
| Use BSNet pseudo-labels with ISBNet instead of GaPro pseudo-labels | ISBNet · S3DIS Area 5 | mAP: 50.5 -> 53.0 | Helped |
| Add local-structure attention (ablation) | SAFormer · ScanNetV2 training overlapping areas | mAcc: 41.5 -> 48.1 | Helped |
| Add global-context attention (ablation) | SAFormer · ScanNetV2 training overlapping areas | mAcc: 41.5 -> 43.5 | Helped |
| Add global-context attention alongside local-structure attention (ablation) | SAFormer · ScanNetV2 training overlapping areas | mAcc: 48.1 -> 52.5 | Helped |
| Add Mean Teacher (ablation) | SAFormer with local/global attention · ScanNetV2 training overlapping areas | mAcc: 52.5 -> 55.3 | Helped |
| Add simulated sample generation (ablation) | SAFormer with Mean Teacher and local/global attention · ScanNetV2 training overlapping areas | mAcc: 55.3 -> 59.6 | Helped |
| Use SAFormer pseudo-labels rather than ignoring overlapping areas (ablation) | ISBNet · ScanNetV2 validation | mAP: 38.1 -> 52.3 | Helped |
| Add soft loss when training with SAFormer pseudo-labels (ablation) | ISBNet · ScanNetV2 validation | mAP: 52.3 -> 52.8 | Helped |
| Use Mean Teacher instead of direct pseudo-label use or iterative self-training (ablation) | SAFormer · ScanNetV2 training overlapping areas | mAcc: 52.5; 52.6; 55.3 | Helped |
| Add gravity/collision constraints when generating simulated samples (ablation) | SAFormer · ScanNetV2 training overlapping areas | mAcc: 58.5 -> 59.3 | Helped |
| Add background points to simulated samples (ablation) | SAFormer · ScanNetV2 training overlapping areas | mAcc: 59.3 -> 59.6 | Helped |
Surprising findings
- Iterative offline self-training barely improves pseudo-label mAcc (52.5 to 52.6), while Mean Teacher raises it to 55.3.
Limitations
- Simulated overlapping samples use simplified translation along X or Y axes and may not match real scene geometry.
- Pseudo-labeler adds about 2.4M parameters and depends on a fixed confidence threshold tau=0.9 and hyperparameters tuned on the data.
Experimental setup
Datasets: ScanNetV2 S3DIS
Benchmarks: ScanNetV2 S3DIS
Metrics: mAP AP@50 AP@25 mAcc
Baselines: Box2Mask WISGP Gapro Mask3D PointGroup SSTNet SoftGroup ISBNet SPFormer CSC PointContrast
Builds on: Mean Teacher SPFormer Gapro Box2Mask ISBNet
Closest papers by meaning
- Sketchy Bounding-box Supervision for 3D Instance Segmentation (CVPR 2025 · similarity 0.91)
- Extreme Point Supervised Instance Segmentation (CVPR 2024 · similarity 0.86)
- MWSIS: Multimodal Weakly Supervised Instance Segmentation with 2D Box Annotations for Autonomous Driving (AAAI 2024 · similarity 0.86)
- SA3DIP: Segment Any 3D Instance with Potential 3D Priors (NeurIPS 2024 · similarity 0.85)
- Insightful Instance Features for 3D Instance Segmentation (CVPR 2025 · similarity 0.85)
- Open3DIS: Open-Vocabulary 3D Instance Segmentation with 2D Mask Guidance (CVPR 2024 · similarity 0.84)
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.