Onset Benchmark 0.1.0rc2

Onset Benchmark 0.1.0rc2

Run your model on standard video and score its physical-event predictions.
This release includes the evaluator, a model-adapter workflow, public labels,
40 prepared REASSEMBLE wrist-camera recordings with frame timestamps, and the
published result tables.

Download the code and videos, then extract both into the same directory.
Use Python 3.11 or later. From the extracted onsetbench directory:

python -m pip install .
python -m onsetbench run --dataset data/reassemble_public40.json --inputs ../inputs/inputs.json --predictor my_adapter:predict --name "My model" --out my-run

Add my_adapter.py in that directory using the adapter instructions.
The runner saves predictions and scores to my-run.

The 40 public recordings are for development. Published results use separate
evaluation sets. The prepared videos retain the original MP4 bytes and native
frame clocks; all 358,330 frames matched their timestamp counts.

• Benchmark code: 342 KB, Apache 2.0.
• Prepared videos and timestamps: 749 MB, CC BY 4.0.
• Published results and documentation: 41 KB.

Use SHA256SUMS to check all three downloads. Source commit:
c7b89e53d5782b355b313fc30f7bfa59ad821614.

Website · Downloads and setup