Research map / Detection and segmentation / CNN features
Scale Optimization Using Evolutionary Reinforcement Learning for Object Detection on Drone Imagery
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 an evolutionary reinforcement learning (EVORL) agent that optimizes per-region scaling factors within a YOLOX-based coarse-to-fine detector for drone imagery, with localization, labeling and scale-consistency rewards and a spatial-semantic attention module.
Problem
Object detection in drone/aerial imagery suffers from large scale variations and small objects; fixed or heuristic patch scaling is suboptimal and no ground-truth optimal scales exist.
Approach
Combines PPO with an evolutionary strategy (crossover and mutation over current actions plus historical best solutions) to pick per-cluster-region scales, guided by localization, labeling and a neighbor scale-consistency reward, plus a spatial-semantic attention for patch features.
Key results
- On UAVDT, the proposed method reaches 28.0 AP, 43.8 AP50 and 31.5 AP75, versus 24.6 AP for the previous best UFPMP-Det.
- On VisDrone validation, the method reaches 42.2 AP versus 40.3 for AdaZoom, with 44.5 AP75.
- Ablation on VisDrone: adding PPO, spatial-semantic attention and evolution raises AP from 37.5 (YOLOX) to 42.2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EVORL instead of UFPMP-Det | Drone object detection · UAVDT | AP: 24.6 -> 28.0 | Helped |
| Use EVORL instead of UFPMP-Det | Drone object detection · UAVDT | AP50: 38.7 -> 43.8 | Helped |
| Use EVORL instead of UFPMP-Det | Drone object detection · UAVDT | AP75: 28.0 -> 31.5 | Helped |
| Use EVORL instead of AdaZoom | Drone object detection · VisDrone val | AP: 40.3 -> 42.2 | Helped |
| Use EVORL instead of AdaZoom | Drone object detection · VisDrone val | AP50: 66.9 -> 66.0 | Hurt |
| Use EVORL instead of AdaZoom | Drone object detection · VisDrone val | AP75: 41.8 -> 44.5 | Helped |
| Add PPO scale optimization (ablation) | YOLOX · VisDrone val | AP: 37.5 -> 39.1 | Helped |
| Add spatial-semantic attention (ablation) | YOLOX + PPO · VisDrone val | AP: 39.1 -> 40.7 | Helped |
| Add evolutionary strategy to PPO (ablation) | YOLOX + PPO + spatial-semantic attention · VisDrone val | AP: 40.7 -> 42.2 | Helped |
Surprising findings
- The method yields a slightly lower AP50 than AdaZoom on VisDrone (66.0 vs 66.9), which the authors attribute to ultra-small objects.
Limitations
- YOLOX struggles to detect ultra-small objects in the coarse stage, lowering AP50 on VisDrone.
- Results of compared methods are taken from original papers rather than re-run under identical settings.
Experimental setup
Datasets: UAVDT VisDrone
Benchmarks: UAVDT VisDrone
Metrics: AP AP50 AP75 APS APM APL
Baselines: Faster R-CNN ClusDet DMNet GLSAN AdaZoom Zoom&Reasoning Det UFPMP-Det SAIC-FPN HRDNet
Builds on: YOLOX Proximal Policy Optimization ClusDet UFPMP-Det AdaZoom
Closest papers by meaning
- Open-World Drone Active Tracking with Goal-Centered Rewards (NeurIPS 2025 · similarity 0.82)
- MAT-Agent: Adaptive Multi-Agent Training Optimization (NeurIPS 2025 · similarity 0.82)
- Gen-Searcher: Reinforcing Agentic Search for Image Generation (NeurIPS 2026 · similarity 0.82)
- Self-Prompting Analogical Reasoning for UAV Object Detection (AAAI 2025 · similarity 0.82)
- Progressive Exploration-Conformal Learning for Sparsely Annotated Object Detection in Aerial Images (NeurIPS 2024 · similarity 0.82)
- PointSR: Self-Regularized Point Supervision for Drone-View Object Detection (CVPR 2025 · 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.