Research map / Detection and segmentation / Open-set and few-shot detection
Fine-Grained Perception in Panoramic Scenes: A Novel Task, Dataset, and Method for Object Importance 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.
Introduces fine-grained object importance ranking (FOIR-360) for panoramic 360° scenes, a weakly supervised pseudo-labeling pipeline based on iterative object erasing and local rank aggregation, a PanoRank network, and the 360Rank test set of 500 images.
Problem
Salient object ranking only ranks salient objects and ignores non-salient ones, and fine-grained importance ranking of all objects in panoramic (360°) scenes lacks supervised data and a suitable annotation protocol.
Approach
Generate importance-ranking pseudo-labels without ground truth by projecting object-centered local views, iteratively erasing the top-ranked object with inpainting to reveal lower-ranked ones, aggregating local rankings via pairwise voting, and combining several 2D saliency models; then train a PanoRank network on these pseudo-labels.
Key results
- On 360Rank, the proposed method reaches SA-SOR 0.658 and SOR 0.877 versus the best 2D SOR baseline QAGNet (SA-SOR 0.650), reported as an 11.5% SOR improvement over QAGNet.
- Plugging the iterative erasing method into SOTA 2D SOR models raises SA-SOR by about 4% on average (e.g., Seqrank 0.594 to 0.621).
- Ablation with RINet as baseline shows the erasing and local rank collector give about a 3% performance gain; LAMA inpainting and max-based Sal-to-Rank performed best.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FOIR-360 instead of existing 2D ranking methods | OURS · 360Rank | SOR: 0.650 -> 0.877 | Helped |
| Use FOIR-360 instead of existing 2D ranking methods | OURS · 360Rank | SA-SOR: 0.650 -> 0.658 | Helped |
| Use FOIR-360 instead of existing 2D ranking methods | OURS · ASSR | SOR: 0.888 -> 0.887 | No clear effect |
| Use FOIR-360 instead of existing 2D ranking methods | OURS · SalMon | SOR: 0.819 -> 0.819 | No clear effect |
| Add iterative erasing and local rank collection (ablation) | RInet · 360Rank | SA-SOR: 0.566 -> 0.591 | Helped |
| Add iterative erasing and local rank collection (ablation) | PSR · 360Rank | SA-SOR: 0.578 -> 0.605 | Helped |
| Add iterative erasing and local rank collection (ablation) | Seqrank · 360Rank | SA-SOR: 0.594 -> 0.621 | Helped |
| Aggregate multiple saliency-model pseudo-labels and fine-tune PanoRank (ablation) | OURS · 360Rank | SOR | Helped |
| Use LAMA rather than MAEF for object erasing (ablation) | OURS · 360Rank | SA-SOR: 0.623 -> 0.628 | Helped |
| Use LAMA rather than MAT for object erasing (ablation) | OURS · 360Rank | SA-SOR: 0.625 -> 0.628 | Helped |
| Use maximum rather than sum pixel values for Sal-to-Rank conversion (ablation) | OURS · 360Rank | SA-SOR: 0.616 -> 0.450 | Hurt |
| Use mean pixel values rather than sum for Sal-to-Rank conversion (ablation) | OURS · 360Rank | SA-SOR: 0.616 -> 0.610 | Hurt |
Surprising findings
- Converting saliency maps to importance rankings by taking the maximum pixel value worked best, and the Max conversion was much worse for the mean-based alternatives in SA-SOR (0.450 vs 0.616 for Sum).
Limitations
- 360Rank is small (500 images) and serves only as a test set.
- The pseudo-label pipeline is offline and time-consuming due to many erasing iterations.
- Underperforms QAGNet on the ASSR dataset, attributed to overfitting of tailored SOR methods and the fine-grained difficulty.
Experimental setup
Datasets: PANDORA AOI ODI-SOD PanopticVideo ASSR SalMon
Benchmarks: 360Rank ASSR SalMon
Metrics: SOR SA-SOR MAE
Baselines: IRSR OCOR PSR SeqRank QAGNet CaSNet UNISAL EMLNet TranSalNet TempSal RINet GSGNet
Builds on: Salient Object Ranking SeqRank Mask2Former PanoSwin LaMa
Closest papers by meaning
- SeqRank: Sequential Ranking of Salient Objects (AAAI 2024 · similarity 0.88)
- Language-Guided Salient Object Ranking (CVPR 2025 · similarity 0.86)
- The Devil is in the Fine-Grained Details: Evaluating Open-Vocabulary Object Detectors for Fine-Grained Understanding (CVPR 2024 · similarity 0.86)
- Benchmarking Large Vision-Language Models on Fine-Grained Image Tasks: A Comprehensive Evaluation (ICLR 2026 · similarity 0.85)
- Same or Not? Enhancing Visual Perception in Vision-Language Models (CVPR 2026 · similarity 0.85)
- More than the Sum: Panorama-Language Models for Adverse Omni-Scenes (CVPR 2026 · 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.