Research map / Detection and segmentation / Open-set and few-shot detection
Self-Prompting Analogical Reasoning for UAV Object Detection
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 SPAR, a YOLOv8-based UAV object detector that generates CLIP-based context-aware prompts for objectness score maps and uses graph convolution over category-level text nodes and pixel-level visual nodes for analogical reasoning.
Problem
Object detection in UAV imagery suffers from small object sizes, varying altitudes and dynamic backgrounds, and visual-only relation reasoning fails to capture semantic relations between objects.
Approach
Combines self-prompting (learnable text context fused with image features via multi-head cross-attention to produce context-aware CLIP text prompts and objectness score maps) with a multi-scale graph whose nodes are class-level prompt embeddings and pixel-level features, using graph convolution to transfer information from easy to hard-to-detect objects.
Key results
- On UAVDT, SPAR reaches 30.5 AP, above EVORL (28.0) and TPH-YOLOv5++ (30.1).
- On VisDrone (validation set), SPAR reaches 42.8 AP versus 41.3 for the YOLOv8 baseline.
- Ablation on VisDrone: adding self-prompting raises YOLOv8 from 41.3 to 42.1 AP; adding analogical reasoning without text nodes gives 42.5; full SPAR gives 42.8.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add self-prompting to YOLOv8 (ablation) | YOLOv8 · VisDrone | AP: 41.3 -> 42.1 | Helped |
| add self-prompting to YOLOv8 (ablation) | YOLOv8 · VisDrone | AP50: 63.3 -> 64.0 | Helped |
| add self-prompting to YOLOv8 (ablation) | YOLOv8 · VisDrone | AP75: 40.9 -> 41.7 | Helped |
| add analogical reasoning without category-level text nodes to +SP (ablation) | +SP · VisDrone | AP: 42.1 -> 42.5 | Helped |
| add analogical reasoning without category-level text nodes to +SP (ablation) | +SP · VisDrone | AP50: 64.0 -> 65.2 | Helped |
| add analogical reasoning without category-level text nodes to +SP (ablation) | +SP · VisDrone | AP75: 41.7 -> 43.6 | Helped |
| add category-level text nodes to analogical reasoning (ablation) | +SP+AR without text node · VisDrone | AP: 42.5 -> 42.8 | Helped |
| add category-level text nodes to analogical reasoning (ablation) | +SP+AR without text node · VisDrone | AP50: 65.2 -> 66.7 | Helped |
| add category-level text nodes to analogical reasoning (ablation) | +SP+AR without text node · VisDrone | AP75: 43.6 -> 45.1 | Helped |
| use SPAR instead of EVORL | UAVDT · UAVDT | AP: 28.0 -> 30.5 | Helped |
| use SPAR instead of EVORL | VisDrone · VisDrone | AP: 42.2 -> 42.8 | Helped |
| use SPAR instead of EVORL | UAVDT · UAVDT small objects | AP S: 21.8 -> 22.9 | Helped |
Surprising findings
- Gains on VisDrone are smaller than on UAVDT, which the authors attribute to semantically similar and visually confusing categories (e.g., pedestrian vs. person).
Limitations
- Gains on VisDrone are modest, especially for fine-grained or semantically confusable categories.
- Pixel-level node selection relies on an empirically set threshold τ.
Experimental setup
Datasets: VisDrone UAVDT
Benchmarks: VisDrone UAVDT
Metrics: AP AP50 AP75 AP_S AP_M AP_L
Baselines: YOLOv8 Faster R-CNN CascadeRCNN ClusDet DMNet TPH-YOLOv5 UFPMP-Det AdaZoom EVORL TPH-YOLOv5++
Builds on: YOLOv8 CLIP CoOp CBAM Chen et al. 2019 graph-based reasoning
Closest papers by meaning
- CP-DETR: Concept Prompt Guide DETR Toward Stronger Universal Object Detection (AAAI 2025 · similarity 0.84)
- PointSR: Self-Regularized Point Supervision for Drone-View Object Detection (CVPR 2025 · similarity 0.84)
- PixDLM: A Dual-Path Multimodal Language Model for UAV Reasoning Segmentation (CVPR 2026 · similarity 0.84)
- A2Seek: Towards Reasoning-Centric Benchmark for Aerial Anomaly Understanding (NeurIPS 2025 · similarity 0.84)
- OVOD-Agent: A Markov-Bandit Framework for Proactive Visual Reasoning and Self-Evolving Detection (CVPR 2026 · similarity 0.83)
- DoGA: Enhancing Grounded Object Detection via Grouped Pre-Training with Attributes (AAAI 2025 · 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.