AeroEyes — AI-Powered Drones for Search & Rescue
Ranked Top 2 nationally · Few-shot CV detecting novel objects from 3 reference images, built for real drone SAR missions
Dec 2025
Top 2
Competition Rank
0.77
Public Score
3 / class
Ref. Images
Problem
Field SAR teams need AI that can generalize to unseen object classes mid-mission — victims, debris, equipment — using drone footage and only 3 labelled examples per class. Standard detection models require thousands of annotations; the challenge is building a system robust enough to rank on a national leaderboard while handling real-world aerial noise, scale variation, and truly novel categories.
Approach
Engineered a few-shot detection pipeline on YOLOE, enabling generalization to completely novel object classes from only 3 reference images — eliminating the need for large per-class annotation sets.
Accelerated class adaptation via prompt-embedding fine-tuning: froze the YOLOE backbone and trained only the detection head and class embeddings — preserving pre-trained representations while cutting overfitting on scarce labels.
Diagnosed and resolved failure modes on small, noisy aerial targets by applying input-level denoising, adaptive resizing, and multi-scale inference — driving a measurable reduction in false positives before final leaderboard submission.
Results
Ranked Top 2 nationally in ZaloAI Challenge 2025 — highest-ranked submission in the Computer Vision track.
Achieved 0.77 public score while generalizing to novel object classes with near-zero labelling effort (3 images per class).
Cut false positives on small aerial targets through multi-scale inference — a key differentiator in the final leaderboard ranking.
Demonstrated that few-shot CV systems can achieve near human-level SAR detection performance without large annotated datasets.
Architecture Overview
3 reference images
Prompt-embedding extraction (YOLOE encoder)
Frozen backbone + trainable detection head/embeddings
Multi-scale inference
NMS
Detection output
How to Run
Clone the repository and install dependencies: pip install -r requirements.txtPrepare reference images (3 per novel class) in /data/reference/Run prompt extraction: python extract_prompts.py --ref-dir data/reference/Run inference: python infer.py --source <drone_footage> --prompts prompts.ptResults saved to /outputs/ with bounding box visualizations