Research map / Point clouds / Segmentation and detection
SASep: Saliency-Aware Structured Separation of Geometry and Feature for Open Set Learning on 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 SASep, which uses Grad-CAM saliency to split point clouds into important and unimportant parts, synthesizes pseudo-unknown objects from unimportant parts, and adds feature-level triplet margin separation for 3D open-set recognition.
Problem
Open-set recognition on 3D point clouds relies on whole-object global features and ignores that different object parts contribute unequally to classification, making known/unknown separation hard.
Approach
Decomposes objects by Grad-CAM saliency into high- and low-saliency parts; re-uses high-saliency parts as extra known-class samples, mixes low-saliency parts into pseudo-unknown objects with label smoothing, and applies a synthetic-feature-aided triplet margin loss on prototype-based cosine logits.
Key results
- On DGCNN synthetic benchmark, SASep reaches AUROC 92.5 and FPR95 28.7 versus best baseline ReAct-level ~88.5 AUROC.
- Ablation on SN2 with DGCNN: full SASep achieves AUROC 94.8, FPR95 24.0, ACC 90.5, mACC 79.8.
- Closed-set ACC on DGCNN improves on SR1 (81.3 vs 72.1) and RR1 (94.0 vs 91.5), with slight drop on SN1 (83.3 vs 84.7).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SASep instead of the strongest listed baseline | DGCNN · Synthetic benchmark | AUROC: 89.3 -> 92.5 | Helped |
| Use SASep instead of the strongest listed baseline | DGCNN · Synthetic-to-Real benchmark | AUROC: 71.6 -> 77.9 | Helped |
| Use SASep instead of the strongest listed baseline | DGCNN · Real-to-Real benchmark | AUROC: 78.0 -> 82.0 | Helped |
| Use SASep instead of the strongest listed baseline | PointNet++ · Synthetic benchmark | AUROC: 88.6 -> 92.8 | Helped |
| Use SASep instead of the strongest listed baseline | PointNet++ · Synthetic-to-Real benchmark | AUROC: 78.2 -> 79.7 | Helped |
| Use SASep instead of the strongest listed baseline | PointNet++ · Real-to-Real benchmark | AUROC: 81.4 -> 84.8 | Helped |
| Add TSD while retaining GSS and SMS (ablation) | SASep · SN2 synthetic benchmark | AUROC: 93.7 -> 94.8 | Helped |
| Add GSS while retaining TSD and SMS (ablation) | SASep · SN2 synthetic benchmark | AUROC: 94.1 -> 94.8 | Helped |
| Add SMS while retaining TSD and GSS (ablation) | SASep · SN2 synthetic benchmark | AUROC: 93.8 -> 94.8 | Helped |
| Add GSS and SMS to TSD (ablation) | SASep · SN2 synthetic benchmark | AUROC: 94.2 -> 94.8 | Helped |
| Add TSD and SMS to GSS (ablation) | SASep · SN2 synthetic benchmark | AUROC: 93.6 -> 94.8 | Helped |
| Add TSD and GSS to SMS (ablation) | SASep · SN2 synthetic benchmark | AUROC: 93.0 -> 94.8 | Helped |
Surprising findings
- Closed-set classification with PointNet++ shows no gain, comparable or slightly lower than the closed-set baseline.
- Under severe corruption the saliency-based decomposition is limited, as the authors acknowledge.
Limitations
- Performance degrades under severe corruption where semantic saliency is hard to capture.
- Saliency maps depend on a pre-trained closed-set backbone computed with Grad-CAM.
- Closed-set accuracy on PointNet++ does not improve over the baseline.
Experimental setup
Datasets: ShapeNetCore ModelNet40 ScanObjectNN
Benchmarks: 3DOS synthetic benchmark (SN1-SN3) 3DOS synthetic-to-real (SR1-SR2) 3DOS real-to-real (RR1-RR3)
Metrics: AUROC FPR95 ACC mACC
Baselines: MSP MLS ODIN Energy GradNorm ReAct VAE NF OE+mixup PointCaM ARPL+CS Cosine proto
Builds on: Grad-CAM 3DOS benchmark maximum logit score prototype-based classifier
Closest papers by meaning
- 3DOS: Towards 3D Open Set Learning - Benchmarking and Understanding Semantic Novelty Detection on Point Clouds (NeurIPS 2022 · similarity 0.84)
- Geometry-Aware 3D Salient Object Detection Network (AAAI 2025 · similarity 0.83)
- OpenNeRF: Open Set 3D Neural Scene Segmentation with Pixel-Wise Features and Rendered Novel Views (ICLR 2024 · similarity 0.83)
- SSLFusion: Scale and Space Aligned Latent Fusion Model for Multimodal 3D Object Detection (AAAI 2025 · similarity 0.82)
- Towards 3D Objectness Learning in an Open World (NeurIPS 2025 · similarity 0.82)
- All Beings Are Equal in Open Set Recognition (AAAI 2024 · similarity 0.82)
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.