Research map / Video / Video segmentation and tracking
CYCLO: Cyclic Graph Transformer Approach to Multi-Object Relationship Modeling in Aerial Videos
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 the AeroEye drone video scene graph dataset and the Cyclic Graph Transformer (CYCLO), which uses cyclic temporal attention over frame-level relation features to model long-range and periodic object relationships.
Problem
Video scene graph generation in drone/aerial videos, where objects are small and dense and existing VidSGG methods struggle with long-term temporal dependencies and lack aerial relationship datasets.
Approach
A cyclic attention mechanism that attends to frame histories with circular index shifts (mod T), so the first and last frames are connected and temporal order is preserved, applied within a multi-head graph transformer on top of DETR-based spatial relation features.
Key results
- On AeroEye PredCls, CYCLO reaches R@20 56.20 and mR@20 19.23, versus 54.18 / 18.37 for HIG.
- On PVSG, CYCLO achieves R/mR@20 of 5.83 / 1.98, versus 4.60 / 1.89 for HIG.
- On ASPIRe relation interactivity, CYCLO gets R@20 15.29 versus 10.26 for HIG.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use cyclic temporal graph attention rather than HIG | VidSGG models · AeroEye PredCls | R@20: 54.18 ± 1.23 -> 56.20 ± 0.70 | Helped |
| use cyclic temporal graph attention rather than HIG | VidSGG models · AeroEye SGCls | mR@20: 15.10 ± 0.37 -> 16.22 ± 0.04 | Helped |
| use cyclic temporal graph attention rather than HIG | VidSGG models · PVSG | R@20: 4.60 -> 5.83 | Helped |
| use cyclic temporal graph attention rather than HIG | VidSGG models · ASPIRe Relation | R@20: 10.26 -> 15.29 | Helped |
| use cyclic attention shift η=1 rather than η=2 (ablation) | CYCLO · AeroEye PredCls | R@20: 55.01 -> 56.20 | Helped |
| use cyclic attention shift η=1 rather than η=3 (ablation) | CYCLO · AeroEye PredCls | R@20: 54.12 -> 56.20 | Helped |
| use cyclic attention shift η=1 rather than η=4 (ablation) | CYCLO · AeroEye PredCls | R@20: 54.55 -> 56.20 | Helped |
| use cyclic attention shift η=1 rather than η=5 (ablation) | CYCLO · AeroEye PredCls | R@20: 54.33 -> 56.20 | Helped |
| discard 2 frames per video (ablation) | CYCLO · AeroEye PredCls | R@20: 56.20 -> 55.08 | Hurt |
| discard 3 frames per video (ablation) | CYCLO · AeroEye PredCls | R@20: 56.20 -> 53.98 | Hurt |
| discard 4 frames per video (ablation) | CYCLO · AeroEye PredCls | R@20: 56.20 -> 52.90 | Hurt |
| discard 5 frames per video (ablation) | CYCLO · AeroEye PredCls | R@20: 56.20 -> 51.84 | Hurt |
Surprising findings
- The paper reports that CYCLO's R/mR ratio on AeroEye is higher than on PVSG, which it interprets as AeroEye being less visually ambiguous.
Limitations
- Performance relies on video continuity; missing or discontinuous frames degrade the cyclic references.
- Absolute mR values on AeroEye remain low (around 13-21%), indicating difficulty with long-tail predicates.
- Frame-discard and shift-value experiments show sensitivity to the temporal ordering and sampling.
Experimental setup
Datasets: AeroEye MAVREC ERA PVSG ASPIRe COCO Visual Genome
Benchmarks: AeroEye PVSG ASPIRe
Metrics: Recall@K mean Recall@K
Baselines: Vanilla Transformer HIG Progression Batch-progression TRACE CSTTran
Builds on: DETR DINO HIG GPT4RoI
Closest papers by meaning
- HyperGLM: HyperGraph for Video Scene Graph Generation and Anticipation (CVPR 2025 · similarity 0.83)
- DSGG: Dense Relation Transformer for an End-to-end Scene Graph Generation (CVPR 2024 · similarity 0.83)
- EGTR: Extracting Graph from Transformer for Scene Graph Generation (CVPR 2024 · similarity 0.82)
- Adaptive Visual Scene Understanding: Incremental Scene Graph Generation (NeurIPS 2024 · similarity 0.82)
- A Closer Look at Dynamic Scene Graph Generation in the Era of Multimodal Large Language Models (NeurIPS 2026 · similarity 0.82)
- OED: Towards One-stage End-to-End Dynamic Scene Graph Generation (CVPR 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.