KissTheHabit/yolov8n-hituav-thermal-finetune

YOLOv8n — HIT-UAV Thermal Fine-Tune

A real, end-to-end weight-updated fine-tune of Ultralytics yolov8n.pt on the HIT-UAV infrared thermal dataset — not prompt-tuning, not a confidence- threshold adjustment. Produced as part of real evidence-gathering for a semantic-ISR research effort; released open because the base YOLOv8 architecture is already open-weight.

Why this exists

The stock yolov8n.pt (COCO-pretrained, never seen thermal/IR data) produces confidently wrong classes on real UAV thermal footage — classes like airplane, scissors, refrigerator, traffic light on scenes that actually contain people, cars, and bicycles. Raising the confidence threshold and swapping to a different COCO-pretrained architecture both failed to fix this. This model fixes it by fine-tuning on real thermal-domain data instead.

Training data and setup

• Dataset: HIT-UAV Infrared Thermal Dataset (CC0 1.0 Universal), official release v1.2.1. COCO-format annotations converted to YOLO format. 5 classes: Person, Car, Bicycle, OtherVehicle, DontCare.
• Split: train 2,029 images / 17,628 boxes, val 290 images / 2,460 boxes, test 579 images / 4,811 boxes — the dataset's own predefined split, untouched.
• Training: yolo detect train model=yolov8n.pt data=data.yaml epochs=100 imgsz=640 device=0. Real GPU (RTX 5080, Blackwell, torch==2.11.0+cu128, torchvision==0.26.0+cu128).
• Weights: best checkpoint by validation mAP50.

Held-out test-split accuracy (real, never seen during training)

Evaluated on the dataset's own 579-image held-out test split:

Class	Images	Instances	Precision	Recall	mAP50	mAP50-95
all	579	4811	0.882	0.763	0.825	0.538
Person	355	2611	0.900	0.872	0.921	0.496
Car	267	1339	0.936	0.963	0.978	0.735
Bicycle	86	796	0.915	0.829	0.910	0.572
OtherVehicle	21	34	0.798	0.697	0.690	0.534
DontCare	23	31	0.862	0.452	0.626	0.354

Person/Car/Bicycle — the classes with the most training instances — are strong (mAP50 0.91–0.98). OtherVehicle/DontCare are weaker, consistent with far fewer training instances (34 and 31 respectively vs. thousands for the top three classes), not a surprising failure mode.

Real before/after result on staged thermal clips

Re-running the same three staged UAV thermal clips through this model instead of stock yolov8n.pt, at the same confidence threshold:

Clip	Stock yolov8n.pt (COCO)	This fine-tune
60m, oblique	person + implausible tail: bird, bench, boat, kite, sheep, train, truck (mean conf. 0.37)	Person, DontCare, Car, OtherVehicle — no implausible classes (mean conf. 0.69)
70m/90°, nadir (worst case)	airplane (55) + scissors (38) — zero plausible classes	Person, Bicycle, Car — fully plausible (mean conf. 0.66)
120m, 30°	traffic light, car, person, truck, refrigerator, boat	Car, Person, OtherVehicle, Bicycle — no implausible classes (mean conf. 0.65)

The nadir clip — the single worst case before fine-tuning, 100% implausible classes — now produces only plausible UAV-thermal classes.

What this establishes and does not

• Establishes: fine-tuning on real thermal-domain data eliminates the implausible-class/domain-shift failure mode on these clips by construction (the model's output vocabulary no longer contains COCO-only classes), and real held-out test-split accuracy (mAP50 0.825 overall) on genuinely unseen data.
• Does not establish: per-instance detection correctness on the three staged clips specifically — no ground-truth labels are staged for them, so precision/recall cannot be computed against them directly. The held-out test-split table above is the real quantitative accuracy evidence; the staged-clip comparison demonstrates the domain-shift fix, a narrower and different claim.
• Does not establish: field-representativeness for any specific operational platform or mission — HIT-UAV is a public academic thermal dataset from a different UAV platform than any specific deployment target.

License

Base model: Ultralytics YOLOv8n (AGPL-3.0). Training data: HIT-UAV Infrared Thermal Dataset (CC0 1.0 Universal). These weights are released under AGPL-3.0, consistent with the base model's license.