pollen-robotics/microduck-duck-detector-dataset

Microducks, seen by a Microduck

Frames from a Microduck's own camera — a wide lens 20 cm off the floor, turned upright by the robot before anything sees it — with bounding boxes on the other Microducks in view. One class, duck. Captured, labelled and trained with pollen-robotics/duck_detector; the model that comes out of it is pollen-robotics/microduck-duck-detector.

Layout

raw/<session>/frame_00000.jpg …   session.json    what the camera saw, and the provenance
labelled/<session>/frame_*.txt    labels.json     the pre-labeller's boxes (Grounding DINO)
reviewed/<session>/frame_*.txt                    what a person corrected — the training labels

Labels are YOLO format: class cx cy w h, normalised. An empty file is a frame with no duck in it, kept on purpose — a detector that has never seen an empty room finds ducks in curtains.

Sessions are the unit. Frames arrive at 2 Hz, so consecutive frames are near-copies; anything that splits this data splits it by session, never by frame.

Sessions

session	tag	robot	frames	Hz	reviewed	note
20260826T120316Z_test1_graphite	test1	graphite	99	2.0	yes
20260826T130804Z_test2_graphite	test2	graphite	121	2.0	yes
20260826T131044Z_test3_graphite	test3	graphite	121	2.0	yes
20260910T152049Z_office-afternoon_graphite	office-afternoon	graphite	60	2.0	yes
20260910T154138Z_office-afternoon_graphite	office-afternoon	graphite	60	2.0	yes

uv run dataset pull          # everything, into datasets/
uv run dataset push          # every local session the Hub does not have yet