Research map / Detection and segmentation / CNN features
SeqRank: Sequential Ranking of Salient Objects
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 SeqRank, a query-based salient object ranking model that predicts attention shifts sequentially via a Fovea Module (high-acuity instance segmentation) and a Sequential Ranking Module (next-fixation selection conditioned on visiting history).
Problem
Salient object ranking predicts the order of human attention over objects in a scene, but existing methods rank all salient objects simultaneously, which does not match human sequential viewing and can yield incorrect attention shift predictions.
Approach
Model foveal vision (RoI-aware fovea layers that refine object queries for high-acuity instance masks) and peripheral vision (a sequential ranking module that picks the next object given the visiting history), invoking the ranker iteratively to produce an ordered list of salient instances.
Key results
- SeqRank reaches SA-SOR 0.685 on ASSR and 0.576 on IRSR, versus 0.650 and 0.543 for the best prior SOR method IRSR, a 6.1% relative SA-SOR gain on IRSR.
- On IRSR, MAE drops to 6.20 from 7.79 for IRSR baseline (about 13% lower).
- Sequential ranking outperforms parallel ranking for both the baseline and SRM variants on ASSR (SA-SOR 0.685 vs 0.676 with SRM).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SeqRank instead of the strongest listed baseline for SA-SOR | SOR methods · ASSR Test Set | SA-SOR: 0.650 -> 0.685 | Helped |
| Use SeqRank instead of the strongest listed baseline for SOR | SOR methods · ASSR Test Set | SOR: 0.873 -> 0.870 | Hurt |
| Use SeqRank instead of the strongest listed baseline for MAE | SOR methods · ASSR Test Set | MAE: 7.31 -> 7.22 | Helped |
| Use SeqRank instead of the strongest listed baseline for SA-SOR | SOR methods · IRSR Test Set | SA-SOR: 0.543 -> 0.576 | Helped |
| Use SeqRank instead of the strongest listed baseline for SOR | SOR methods · IRSR Test Set | SOR: 0.820 -> 0.822 | Helped |
| Use SeqRank instead of the strongest listed baseline for MAE | SOR methods · IRSR Test Set | MAE: 7.14 -> 6.20 | Helped |
| Replace the transformer decoder layer with the fovea layer (ablation) | SeqRank on ASSR · ASSR | SA-SOR: 0.672 -> 0.685 | Helped |
| Replace the transformer decoder layer with the fovea layer (ablation) | SeqRank on ASSR · ASSR | MAE: 7.82 -> 7.22 | Helped |
| Use RoI window sizes h=w=1/2/3 rather than h=w=1/1/1 (ablation) | SeqRank on ASSR · ASSR | MAE: 7.35 -> 7.22 | Helped |
| Use multi-scale features instead of feat5 only (ablation) | SeqRank on ASSR · ASSR | SA-SOR: 0.676 -> 0.685 | Helped |
| Use multi-scale features instead of feat3 only (ablation) | SeqRank on ASSR · ASSR | MAE: 7.44 -> 7.22 | Helped |
| Use 6 rather than 3 fovea layers with multi-scale features (ablation) | SeqRank on ASSR · ASSR | MAE: 7.59 -> 7.22 | Helped |
Surprising findings
- The standard transformer decoder layer and the fovea layer differ modestly, but the sequential strategy gains are smaller than expected in SOR metric for the baseline (0.870 vs 0.861 parallel).
Limitations
- Gains on SOR metric over strongest baselines are small (e.g., 0.870 vs 0.873 OCOR on ASSR); SA-SOR is the main improvement.
- Evaluated only on two SOR benchmarks (ASSR, IRSR) with a limited set of scenes.
- Sequential inference runs the ranking module repeatedly up to N=100 times, which may add latency.
Experimental setup
Datasets: ASSR IRSR MS-COCO SALICON ImageNet
Benchmarks: ASSR IRSR
Metrics: SA-SOR SOR MAE
Baselines: VST MENet S4Net QueryInst Mask2Former RSDNet ASRNet PPA IRSR OCOR
Builds on: Mask2Former FPN Swin Transformer DETR-style Hungarian matching
Closest papers by meaning
- Language-Guided Salient Object Ranking (CVPR 2025 · similarity 0.91)
- Fine-Grained Perception in Panoramic Scenes: A Novel Task, Dataset, and Method for Object Importance Ranking (AAAI 2025 · similarity 0.88)
- Advancing Saliency Ranking with Human Fixations: Dataset Models and Benchmarks (CVPR 2024 · similarity 0.86)
- CaRDiff: Video Salient Object Ranking Chain of Thought Reasoning for Saliency Prediction with Diffusion (AAAI 2025 · similarity 0.86)
- A Motion-aware Spatio-temporal Graph for Video Salient Object Ranking (NeurIPS 2024 · similarity 0.85)
- The All-Seeing Project: Towards Panoptic Visual Recognition and Understanding of the Open World (ICLR 2024 · 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.