Research map / Point clouds / Segmentation and detection
MSV-PCT: Multi-Sparse-View Enhanced Transformer Framework for Salient Object Detection in Point Clouds
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 MSV-PCT, a point cloud salient object detection framework that fuses a PSOD-Net backbone with three MLP-based orthographic multi-view branches and an edge-weighted cross-entropy loss, and builds SOD-RGBD2PC point cloud datasets from RGB-D data.
Problem
Salient object detection on 3D point clouds is limited; prior methods treat it as point cloud segmentation over all points, ignoring human surface-only (occlusion-limited) perception, and suffer from edge artifacts and few datasets.
Approach
Augment a point cloud transformer backbone with lightweight PointMLP branches operating on front, side, and top orthographic projections of the point cloud (background points removed via a depth-threshold filter), plus a point-cloud edge-aware weighted cross-entropy loss.
Key results
- On PCSOD, MSV-PCT reaches IoU 0.763 and F-measure 0.845 vs. PSOD-Net's 0.711 and 0.805 (about 5% IoU gain).
- On the converted NPLR dataset, MSV-PCT achieves MAE 0.056 and IoU 0.760, beating PSOD-Net (MAE 0.063, IoU 0.707).
- Ablation on PCSOD shows adding all three views (XY+YZ+XZ) gives IoU 0.763 vs. 0.719 for a single front view; the edge loss raises IoU from 0.736 to 0.763.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MSV-PCT instead of PSOD-Net | Point cloud salient object detection · NPLR | F-measure: 0.800 -> 0.846 | Helped |
| use MSV-PCT instead of PSOD-Net | Point cloud salient object detection · RGB-D Scenes v2 | F-measure: 0.805 -> 0.850 | Helped |
| use MSV-PCT instead of PSOD-Net | Point cloud salient object detection · PCSOD | IoU: 0.711 -> 0.763 | Helped |
| use all three views (XY+YZ+XZ) instead of only XY (ablation) | MSV-PCT · PCSOD | IoU: 0.719 -> 0.763 | Helped |
| use all three views (XY+YZ+XZ) instead of only YZ (ablation) | MSV-PCT · PCSOD | IoU: 0.698 -> 0.763 | Helped |
| use all three views (XY+YZ+XZ) instead of only XZ (ablation) | MSV-PCT · PCSOD | IoU: 0.715 -> 0.763 | Helped |
| use all three views (XY+YZ+XZ) instead of XY+YZ (ablation) | MSV-PCT · PCSOD | IoU: 0.702 -> 0.763 | Helped |
| use all three views (XY+YZ+XZ) instead of XY+XZ (ablation) | MSV-PCT · PCSOD | IoU: 0.730 -> 0.763 | Helped |
| use all three views (XY+YZ+XZ) instead of YZ+XZ (ablation) | MSV-PCT · PCSOD | IoU: 0.724 -> 0.763 | Helped |
| add weighted cross-entropy edge loss (ablation) | MSV-PCT · PCSOD | IoU: 0.736 -> 0.763 | Helped |
Surprising findings
- The paper reports that the single-view ablation shows the top view (YZ) alone performs worse than the front view, yet combining views consistently improves results.
Limitations
- Converted SOD-RGBD2PC labels are pseudo ground truth produced by D3Net rather than human annotations.
- Only three fixed orthographic views are explored; no analysis of other view configurations beyond the ablation.
- No variance or significance statistics reported for the results.
Experimental setup
Datasets: PCSOD NPLR RGB-D Scenes v2
Benchmarks: PCSOD NPLR (SOD-RGBD2PC) RGB-D Scenes v2 (SOD-RGBD2PC)
Metrics: MAE F-measure E-measure IoU
Baselines: ASSANet PointTransformer PCT PointMLP PointNeXt ShellNet PointSal PSOD-Net
Builds on: PSOD-Net PointMLP PointSal D3Net
Closest papers by meaning
- Geometry-Aware 3D Salient Object Detection Network (AAAI 2025 · similarity 0.89)
- WeakPCSOD: Overcoming the Bias of Box Annotations for Weakly Supervised Point Cloud Salient Object Detection (AAAI 2024 · similarity 0.87)
- One for All: Multi-Domain Joint Training for Point Cloud Based 3D Object Detection (NeurIPS 2024 · similarity 0.85)
- ImOV3D: Learning Open Vocabulary Point Clouds 3D Object Detection from Only 2D Images (NeurIPS 2024 · similarity 0.85)
- LION: Linear Group RNN for 3D Object Detection in Point Clouds (NeurIPS 2024 · similarity 0.85)
- DiMSOD: A Diffusion-Based Framework for Multi-Modal Salient Object Detection (AAAI 2025 · similarity 0.85)
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.