ander2221/visdrone-yolo9-preview
YOLOv9-t fine-tuned on VisDrone2019-DET aerial imagery using LibreYOLO. Ten classes (pedestrian, people, bicycle, car, van, truck, tricycle, awning-tricycle, bus, motor), top-down drone perspective.
Companion use case: LibreYOLO/use-cases/visdrone-finetune.
Training
- size:
t - imgsz:
384 - epochs:
5 - dataset: VisDrone2019-DET via Voxel51's HuggingFace mirror
- compute: Apple Metal Performance Shaders (MPS, M-series GPU)
Metrics
{}
Usage — Python
from huggingface_hub import hf_hub_download
from libreyolo import LibreYOLO
ckpt = hf_hub_download(repo_id="ander2221/visdrone-yolo9-preview", filename="visdrone.pt")
model = LibreYOLO(ckpt)
result = model("aerial.jpg")
for box, cls, conf in zip(result.boxes.xyxy, result.boxes.cls, result.boxes.conf):
print(box, ["pedestrian","people","bicycle","car","van","truck","tricycle","awning-tricycle","bus","motor"][int(cls)], float(conf))
Usage — ONNX (browser, edge, cross-runtime)
import onnxruntime as ort
from huggingface_hub import hf_hub_download
onnx = hf_hub_download(repo_id="ander2221/visdrone-yolo9-preview", filename="visdrone.onnx")
session = ort.InferenceSession(onnx, providers=["CPUExecutionProvider"])
# Preprocess image to (1, 3, 384, 384) float32 in [0,1] then:
out = session.run(None, {"images": preprocessed})
A live browser demo using this ONNX is at https://libreyolo.github.io/use-cases/visdrone-finetune/demo/ (zero-install, runs locally in Chrome via WebGPU/onnxruntime-web).
Classes (index → name)
| idx | name |
|---|---|
| 0 | pedestrian |
| 1 | people |
| 2 | bicycle |
| 3 | car |
| 4 | van |
| 5 | truck |
| 6 | tricycle |
| 7 | awning-tricycle |
| 8 | bus |
| 9 | motor |
License
MIT (the model file). Dataset (VisDrone2019-DET) is governed by its own license terms — please review for your use case.