Research map / Detection and segmentation / Open-set and few-shot detection
Domain Separation Graph Neural Networks for Saliency Object Ranking
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 DSGNN, a domain-separation graph neural network for saliency object ranking that builds separate shape and texture object graphs with multi-dimensional edges and a cross-image saliency subspace, reaching SOTA on ASSR and IRSR.
Problem
Saliency object ranking (SOR) needs to rank the saliency degrees of multiple objects in an image, but prior methods do not explicitly model saliency-related relationships between objects' shapes and textures, nor share saliency structure across images, and use single-value graph edges.
Approach
Separately extract shape and texture graphs over object queries, disentangle saliency-relevant from irrelevant subspaces via GNN encoders with difference/similarity/reconstruction losses (STGDS), add a cross-image graph module (CIGDS) for a unified saliency-level representation, and learn multi-dimensional edge features with GGCN.
Key results
- On ASSR with Swin-L, DSGNN reaches MAE 0.047, SOR 0.853, SA-SOR 0.761, versus OCOR's 0.079 / 0.883 / 0.687.
- On IRSR with Swin-L, DSGNN reports SA-SOR 0.606 and MAE 0.067; the ResNet-50 variant gets SA-SOR 0.568.
- Ablation: full STGDS+CIGDS with shape and texture branches improves SA-SOR from 0.739 (shape-only baseline) to 0.761.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DSGNN with Swin-L instead of IRSR | DSGNN · IRSR test | SA-SOR: 0.565 -> 0.606 | Helped |
| Add texture information to shape-only features (ablation) | DSGNN · ASSR ablation | SOR: 0.835 -> 0.836 | Helped |
| Add the STGDS shape branch (ablation) | DSGNN · ASSR ablation | SA-SOR: 0.744 -> 0.755 | Helped |
| Add the STGDS texture branch (ablation) | DSGNN · ASSR ablation | SA-SOR: 0.744 -> 0.758 | Helped |
| Add the CIGDS module (ablation) | DSGNN · ASSR ablation | SA-SOR: 0.759 -> 0.761 | Helped |
| Combine similarity, difference, and reconstruction losses (ablation) | STGDS · ASSR ablation | SA-SOR: 0.753 -> 0.761 | Helped |
| Combine similarity and difference losses (ablation) | STGDS · ASSR ablation | SA-SOR: 0.746 -> 0.758 | Helped |
| Combine similarity and reconstruction losses (ablation) | STGDS · ASSR ablation | SA-SOR: 0.746 -> 0.756 | Helped |
| Combine difference and reconstruction losses (ablation) | STGDS · ASSR ablation | SA-SOR: 0.750 -> 0.753 | Helped |
| Combine similarity and difference losses (ablation) | CIGDS · ASSR ablation | MAE: 0.050 -> 0.047 | Helped |
| Replace MLP with GGCN (ablation) | DSGNN graph modules · ASSR ablation | MAE: 0.052 -> 0.049 | Helped |
| Use multi-dimensional edges instead of standard GGCN edges (ablation) | GGCN · ASSR ablation | MAE: 0.049 -> 0.047 | Helped |
Surprising findings
- The ResNet-50 variant outperforms previous Swin-L-based models on MAE and SA-SOR, and the authors note DSGNN did not achieve SOTA on the SOR metric.
Limitations
- Does not achieve state-of-the-art on the SOR metric, which the authors say ignores missed targets.
- Reproduced OCOR results were not compared because they differ from reported results.
Experimental setup
Datasets: ASSR IRSR MS-COCO SALICON
Benchmarks: ASSR IRSR
Metrics: MAE SOR SA-SOR
Baselines: RSDNet ASSR IRSR SOR OCOR S4Net BASNet CPD-R SCRN
Builds on: Mask2Former GGCN ASSR IRSR Domain Separation Networks
Closest papers by meaning
- Promoting Saliency From Depth: Deep Unsupervised RGB-D Saliency Detection (ICLR 2022 · similarity 0.85)
- Image Saliency Prediction in Transformed Domain: A Deep Complex Neural Network Method (AAAI 2019 · similarity 0.85)
- Language-Guided Salient Object Ranking (CVPR 2025 · similarity 0.84)
- EGTR: Extracting Graph from Transformer for Scene Graph Generation (CVPR 2024 · similarity 0.83)
- Advancing Saliency Ranking with Human Fixations: Dataset Models and Benchmarks (CVPR 2024 · similarity 0.83)
- SeqRank: Sequential Ranking of Salient Objects (AAAI 2024 · similarity 0.83)
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.