NightEye v2 — thermal UAV person detector
YOLO26s thermal person detector for nadir UAV search-and-rescue imagery.
Metrics
| Evaluation | Metric | Result |
|---|---|---|
| Internal val | mAP50 / mAP50-95 | ~0.918 / ~0.556 |
| POP test | mAP50 | 0.136 → 0.522 |
| HIT-UAV test (v1 baseline) | mAP50 | 0.948 |
Caveat: v2 was trained with 261 pseudo-labeled 2026s2 frames. 2026s2
is not a clean test for v2. A clean 2026s2 gold set is being labeled and
v3 is being trained without any 2026s2 data.
Usage
from ultralytics import YOLO
model = YOLO("best.pt")
results = model.predict("thermal.jpg", imgsz=640, conf=0.25)
Training data
- HIT-UAV (CC0)
- TFST thermal (private, consent-bound; not redistributed)
- POP (external; license undocumented)
Limitations
- Small objects at high altitude; occlusion; thermal polarity changes.
- Single-frame detector; no tracker.
- Not flight-qualified.