Research map / Point clouds / Segmentation and detection
Geometry-Aware 3D Salient Object Detection Network
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 3DGAS, a point cloud salient object detection network that clusters points into superpoints and uses superpoint-point attention plus a class-agnostic metric loss to sharpen object boundaries on PCSOD.
Problem
Point cloud salient object detection produces blurry object boundaries in complex backgrounds because existing methods under-use the geometric structure of 3D objects.
Approach
Cluster points into superpoints via a distance- and feature-similarity greedy algorithm, pool them into superpoint features, and use superpoint-point cross-attention to inject geometric structure back into point features, supervised by a class-agnostic pull/push loss on local object regions.
Key results
- On PCSOD with xyz+rgb input, 3DGAS reaches MAE 0.042, F-measure 0.848, E-measure 0.912, IoU 0.763, beating EPFNet (MAE 0.047, IoU 0.727).
- With xyz only, 3DGAS achieves IoU 0.720 versus 0.587 for PSOD-Net re-run with official code.
- Reported inference time 0.06s with 0.2M parameters, the lowest among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use 3DGAS instead of EPFNet | 3D salient object detection · PCSOD test (xyz+rgb) | MAE: 0.047 -> 0.042 | Helped |
| use 3DGAS instead of EPFNet | 3D salient object detection · PCSOD test (xyz+rgb) | F-measure: 0.820 -> 0.848 | Helped |
| use 3DGAS instead of EPFNet | 3D salient object detection · PCSOD test (xyz+rgb) | E-measure: 0.898 -> 0.912 | Helped |
| use 3DGAS instead of EPFNet | 3D salient object detection · PCSOD test (xyz+rgb) | IoU: 0.727 -> 0.763 | Helped |
| use 3DGAS instead of PSOD-Net* | 3D salient object detection · PCSOD test (xyz) | MAE: 0.087 -> 0.050 | Helped |
| add superpoint partition (SP) (ablation) | 3DGAS baseline · PCSOD test | MAE: 0.116 -> 0.050 | Helped |
| add superpoint partition (SP) (ablation) | 3DGAS baseline · PCSOD test | F-measure: 0.632 -> 0.810 | Helped |
| add geometry enhancement (GE) to SP (ablation) | 3DGAS with SP · PCSOD test | MAE: 0.050 -> 0.047 | Helped |
| add geometry enhancement (GE) to SP (ablation) | 3DGAS with SP · PCSOD test | F-measure: 0.810 -> 0.832 | Helped |
| add point cloud class-agnostic loss (CA) to SP + GE (ablation) | 3DGAS with SP + GE · PCSOD test | MAE: 0.047 -> 0.042 | Helped |
| add point cloud class-agnostic loss (CA) to SP + GE (ablation) | 3DGAS with SP + GE · PCSOD test | F-measure: 0.832 -> 0.848 | Helped |
Surprising findings
- Superpoint partition alone raises the ablation baseline IoU from 0.519 to 0.720, the largest single gain in the ablation.
Limitations
- Superpoint generation depends on the quality of learned features and distance/feature thresholds.
- EPFNet comparison uses RGB images, so the single-modal comparison is not like-for-like.
- Higher FLOPs (2.6G) than PointSal (1.4G).
Experimental setup
Datasets: PCSOD
Benchmarks: PCSOD
Metrics: MAE F-measure E-measure IoU
Baselines: Point Transformer PointNeXt PointSal PSOD-Net EPFNet
Builds on: PointSal PSOD-Net EPFNet PointNet++ Point Transformer
Closest papers by meaning
- MSV-PCT: Multi-Sparse-View Enhanced Transformer Framework for Salient Object Detection in Point Clouds (AAAI 2025 · similarity 0.89)
- Density-Based Clustering for 3D Object Detection in Point Clouds (CVPR 2020 · similarity 0.88)
- Locate Globally, Segment Locally: A Progressive Architecture With Knowledge Review Network for Salient Object Detection (AAAI 2021 · similarity 0.87)
- SPGroup3D: Superpoint Grouping Network for Indoor 3D Object Detection (AAAI 2024 · similarity 0.87)
- Structure Aware Single-Stage 3D Object Detection From Point Cloud (CVPR 2020 · similarity 0.87)
- WeakPCSOD: Overcoming the Bias of Box Annotations for Weakly Supervised Point Cloud Salient Object Detection (AAAI 2024 · similarity 0.86)
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.