Research-0082 — BVI-DVC corpus feasibility for fr_regressor_v2¶

Question¶

Can adding the BVI-DVC reference corpus (Bristol VI Lab, 2021) to the fr_regressor_v2 training corpus alongside the Netflix Public drop produce a measurable LOSO PLCC lift, given the license, content-overlap, and fold-expansion posture the fork commits to?

Method¶

1. Audit the BVI-DVC license terms via the Bristol Visual Information Lab portal and Zenodo metadata; classify against fork redistribution rules.
2. Compare BVI-DVC's content distribution against the Netflix Public drop (9 sources) — qualitative content-class overlap (cinematic film_drama, sports, animation, wildlife, urban, texture).
3. Project the LOSO partition expansion: today the trainer holds out one of 9 sources per fold; with BVI-DVC's tier-D added, the partition expands to 9 + N source-folds.

Result¶

1. License analysis¶

BVI-DVC is distributed under research-only terms — the corpus may be used for research and may not be redistributed. The fork's existing posture for the Netflix Public drop (ADR-0203) is the same model: local-only handling, derived weights ship, source corpus does not. The redistribution rule for BVI-DVC therefore requires no new infrastructure; it requires the same gitignore + README discipline already in place for .workingdir2/netflix/.

Concrete rules:

• .workingdir2/BVI-DVC Part 1.zip — gitignored.
• .workingdir2/bvi-dvc-extracted/ — gitignored.
• runs/full_features_bvi_dvc_*.parquet — gitignored.
• runs/bvi_dvc_corpus.jsonl — gitignored.
• ~/.cache/vmaf-tiny-ai-bvi-dvc-full/ — outside repo, never committed.
• model/tiny/fr_regressor_v2*.onnx — derived weights, shippable.

2. Content overlap with Netflix Public¶

The Netflix drop under-represents urban and texture-heavy content; BVI-DVC fills both gaps. Animation has no BVI-DVC analogue, so BigBuckBunny remains the only anchor for that class.

3. LOSO partition expansion¶

Netflix-only LOSO trains 9 models, one per held-out source. Adding BVI-DVC tier-D (~120 sources) expands the partition to 9 + ~120 source-folds. Each fold trains on a much larger remainder set, which should narrow the per-fold variance band.

Trade-off: tier-D's resolution (480 × 272) is well below typical production resolutions, so weighting BVI-DVC tier-D folds equally with Netflix HD/UHD folds may bias the regressor toward low-res behaviour. Two mitigations are available:

• include tier-B (1920 × 1088) and tier-C (960 × 544) instead of / alongside tier-D — increases extraction cost ~16× per clip;
• weight folds by source-resolution category in the LOSO aggregator — changes only the reporting layer.

The ingestion ADR-0310 ships the infrastructure tier-agnostic; tier selection is a runtime flag, deferred to the multi-seed sweep.

Conclusion¶

License is compatible with the fork's existing local-only corpus posture. Content overlap is favourable: BVI-DVC fills the urban and texture gaps the Netflix drop has and reinforces high-motion. The LOSO partition expansion materially widens the training surface. The infrastructure to make this measurable (JSONL adapter + merge utility

• tests) is small and ships with ADR-0310.

The actual PLCC measurement and the production-weights flip are deferred to a multi-seed sweep that runs outside this PR. The flip gate stays anchored on ADR-0303's ensemble criterion; a corpus expansion that does not lift mean LOSO PLCC by ≥ 1σ above the Netflix-only baseline is not shipped to production weights.

References¶

• Ma, Zhang, Bull. BVI-DVC: A Training Database for Deep Video Compression. IEEE Transactions on Multimedia, 2021.
• ADR-0203 — Netflix Public drop redistribution posture.
• ADR-0235 — fr_regressor_v2.
• ADR-0303 — ensemble-flip ship gate (corpus-expansion ship criterion lives here).
• ADR-0310 — this digest's decision record.