Research protocol

The working question is whether detector-agnostic temporal repair can reduce box jitter, class switching, and short gaps without materially degrading task accuracy or adding unacceptable latency.

The repository's research/ directory contains the protocol, artifact analysis, plot generation, citation metadata, hypotheses, threats to validity, and paper outline. Dataset content, model weights, and result artifacts are not bundled.

MOT17 preparation requires an existing user download and exact acknowledgement of its CC BY-NC-SA 3.0 terms:

uv run python scripts/prepare_temporalfix_dataset.py path/to/MOT17 \
  --accept-license CC-BY-NC-SA-3.0 \
  --output benchmark_artifacts/datasets/mot17-manifest.json

The command hashes files into a relative manifest; it does not copy, download, or upload the dataset.