Research map / Detection and segmentation / Weakly supervised detection and segmentation
WeakPCSOD: Overcoming the Bias of Box Annotations for Weakly Supervised Point Cloud Salient Object Detection
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 WeakPCSOD, the first weakly supervised point cloud salient object detection method trained only from 3D bounding boxes, using a mask-to-box transformation, color consistency loss, and self-training.
Problem
Point cloud salient object detection needs costly dense point-level annotations; naive box-based supervision introduces strong box shape bias and noisy labels.
Approach
Supervise the outer box mask of the prediction (via differentiable voxelization, projection and back-projection) instead of the prediction itself, so shape is decoupled from box bias; add a color-consistency loss on high-similarity point pairs and a self-training stage with box-filtered pseudo masks.
Key results
- WeakPCSOD with PointNet++ reaches Fm 0.780 on the test set vs 0.614 for cube-mask supervision, a 16.6% gain.
- WeakPCSOD with PointMLP reaches Fm 0.804, Em 0.896, MAE 0.058, IoU 0.706 on the test set, outperforming the weakly supervised Box2Mask (Fm 0.721) and approaching fully supervised PointSOD (Fm 0.772).
- Ablation with PointNet++: M2B raises Fm from 0.614 to 0.736, CC loss adds 1.2 points, and self-training adds 3.2 points, reaching 0.780.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add M2B transformation (ablation) | WeakPCSOD with PointNet++ · PCSOD test set | Fm: 0.614 -> 0.736 | Helped |
| add color consistency loss (ablation) | WeakPCSOD with PointNet++ and M2B · PCSOD test set | Fm: 0.736 -> 0.748 | Helped |
| add self-training (ablation) | WeakPCSOD with PointNet++ and M2B + CC loss · PCSOD test set | Fm: 0.748 -> 0.780 | Helped |
Surprising findings
- WeakPCSOD with PointMLP (Fm 0.804) exceeds the fully supervised PointSOD baseline (Fm 0.772) on the test set.
Limitations
- Relies on the assumption that points with similar color share the same label, which is not rigorous and introduces noise.
- Mask-to-box transformation can yield non-unique predictions for one outer box mask, requiring the CC loss to disambiguate.
- Evaluated only on the single PointSOD dataset with two backbones.
Experimental setup
Datasets: PointSOD dataset (2,872 indoor/outdoor 3D scenes)
Benchmarks: PointSOD PCSOD testing set (872 scenes)
Metrics: Fm Em MAE IoU
Baselines: PointNet PointNet++ PointCNN ShellNet RandLA PointSOD BCM BBTP Box2Mask
Builds on: PointSOD PointNet++ PointMLP Box2Mask
Closest papers by meaning
- VSRD: Instance-Aware Volumetric Silhouette Rendering for Weakly Supervised 3D Object Detection (CVPR 2024 · similarity 0.88)
- Leveraging Imagery Data with Spatial Point Prior for Weakly Semi-supervised 3D Object Detection (AAAI 2024 · similarity 0.87)
- Multi-Modality Affinity Inference for Weakly Supervised 3D Semantic Segmentation (AAAI 2024 · similarity 0.87)
- Point2RBox: Combine Knowledge from Synthetic Visual Patterns for End-to-end Oriented Object Detection with Single Point Supervision (CVPR 2024 · similarity 0.87)
- MSV-PCT: Multi-Sparse-View Enhanced Transformer Framework for Salient Object Detection in Point Clouds (AAAI 2025 · similarity 0.87)
- ABBSPO: Adaptive Bounding Box Scaling and Symmetric Prior based Orientation Prediction for Detecting Aerial Image Objects (CVPR 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.