Microduck duck detector
Finds other Microducks in a Microduck's camera. One class, duck, 320×320 letterboxed input,
2100 candidate boxes out; trained with
pollen-robotics/duck_detector on
pollen-robotics/microduck-duck-detector-dataset and run on the robot by
duck-detect in pollen-robotics/microduck.
Current run: v2 — duck_detect.pt, duck_detect.onnx, duck_detect.rknn. Every run is a git tag on this repo; main is the latest.
| metric | value |
|---|---|
| map50 | 0.7999 |
| map50_95 | 0.5616 |
| precision | 0.9564 |
| recall | 0.6818 |
| imgsz | 320 |
| model | yolo11n.pt |
| smoke | False |
Trained on these sessions of the dataset, whole sessions held out for validation:
20260826T120316Z_test1_graphite20260826T130804Z_test2_graphite20260826T131044Z_test3_graphite20260910T152049Z_office-afternoon_graphite20260910T154138Z_office-afternoon_graphite(val)
duck_detect.rknn is INT8 for the RK3566's NPU, quantised against real frames; .onnx is the
float model it came from (static shapes, opset 12), and .pt is the ultralytics checkpoint.
The int8 output carries its own scale, so a confidence threshold has to be set against the
RKNN, not inherited from the float model.
uv run model pull # weights/duck_detect.* from main
uv run model pull --revision v2
- Downloads last month
- -