NULLXES CERBER-CV (v1)

Civil aerial scene detector for the NULLXES CERBER perception stack.
YOLOv8s fine-tuned on VisDrone-DET with CERBER class remap.

Metrics (VisDrone val, 548 images)

Class	P	R	mAP50	mAP50-95
all	0.818	0.692	0.760	0.439
human	0.779	0.604	0.672	0.313
vehicle	0.857	0.779	0.848	0.565

Train: imgsz 1280, batch 32, 100 epochs, ~2.0 h on RTX PRO 6000.
Export ONNX: imgsz 640, opset 17.
detector_alpha.onnx sha256: 40151159e7bf59fcfc24b591124ff7aeec88ff4365619ee701fc186efdce199a

Classes (head nc=13)

Trained with labels: human (0), vehicle (1).
Other CERBER ids (uav, landing_zone, … cargo) are in the head but untrained in v1.

Files

• best.pt — Ultralytics weights
• best.onnx / detector_alpha.onnx — flight ONNX (yolo_v8_raw)
• results.csv / results.png — train curves

Cite

VisDrone: https://github.com/VisDrone/VisDrone-Dataset
Ultralytics YOLO (AGPL-3.0)