Research-0046 — Bristol VI-Lab dataset feasibility for tiny-AI training and parity soak¶

Why now¶

fr_regressor_v1 (T6-1a, commit e421d70) is trained on Netflix Public only — 9 reference + 70 distorted, ~37 GB, gitignored at .workingdir2/netflix/. The codec-aware fr_regressor_v2 plan in ADR-0042 and ADR-0235 needs a wider codec sweep with subjective labels. Separately, the cross-backend parity gate (/cross-backend-diff) currently runs over only the 3 Netflix golden pairs at places=4; we have no real-corpus soak.

The Bristol Visual Information Lab (David Bull's group, "VI-Lab") publishes a family of public BVI-* datasets. Several are MOS-labelled and codec-diverse. The user has asked for a feasibility report before any download — idle GPUs are available, but storage and licence posture have to be cleared first.

1. Inventory¶

Sizes are estimates from paper specs (sequence count × resolution × duration × bit-depth × 1.5 chroma factor); ranges where the lab page does not state a download size. Format column: YUV means raw 4:2:0 planar, Y4M means raw with header, MP4 means encoded streams. All Bristol downloads are gated by a registration form (MS OneDrive or data.bris.ac.uk); confirmation typically lands inside 2 days for the OneDrive route.

Notes:

• BVI-CC total comes out at 9 src × 4 resolutions × ~5 s × 60 fps × ~10-bit 4:2:0 plus 306 distorted; the raw-source footprint alone is on the order of 80–150 GB and the distorted leg adds the rest.
• BVI-AOM is the only entry with an authoritative size figure (124 GB before zip) because the GitHub repo states it explicitly.
• BVI-DVC and BVI-AOM are training corpora and ship without subjective labels; they're useful for parity soak and self-supervised feature extraction, not for fr_regressor_v2's labelled-MOS leg.

2. Licence audit¶

Common posture across the BVI- family: research / academic use, registration form before download, redistribution of the raw sequences* is restricted, but derivative products (extracted features, computed metrics, statistics) are generally permissible with attribution. Per-dataset detail:

Specific flags to watch:

1. BVI-AOM CableLabs subset — CC-BY-NC-ND 3.0 means we cannot ship derivative files for those sequences. The published paper marks them; ingest must propagate that flag through to any parquet manifest so per-clip filtering is possible at training time and at parquet-publish time.
2. BVI-DVC "training only" clause — the dataset is licensed for training video coding tools. Using it as input to a tiny-AI quality regressor sits at the edge of that wording. Defensible under "objective metric research" but worth one explicit user confirmation before we bake it into a published model card.
3. No dataset has a clean redistribute raw clips clause. The fork must never check Bristol clips into the repo (.gitignore already covers the Netflix corpus pattern; mirror it for BVI).

3. Use-case fit¶

fr_regressor_v2 cares about (a) codec one-hot, (b) reliable MOS labels, (c) reasonable diversity. BVI-CC's 9×34 = 306 labelled distorted sequences across HM/AV1/VTM at four resolutions hits all three. BVI-HD adds 384 HEVC-distorted sequences with DMOS at HD only — useful as a held-out single-codec validation slice.

4. Effort to extract one dataset (BVI-CC)¶

BVI-CC is the smallest useful MOS-labelled candidate and has the codec axis the v2 regressor needs.

# 1. Submit registration form (manual; ~2 day SLA)
#    https://fan-aaron-zhang.github.io/BVI-CC/

# 2. Stage download (MS OneDrive link arrives by email).
#    Pull into the gitignored corpus root:
mkdir -p .workingdir2/bristol/bvi-cc
rclone copy onedrive:BVI-CC .workingdir2/bristol/bvi-cc \
    --transfers 4 --progress

# 3. Most BVI clips ship as raw YUV with sidecar names that encode
#    geometry (e.g. *_3840x2160_60fps_10bit_420.yuv). No ffmpeg
#    transcode is required to feed libvmaf — vmaf consumes raw YUV
#    directly. If a clip is delivered as Y4M, strip the header:
ffmpeg -i src.y4m -f rawvideo -pix_fmt yuv420p10le src.yuv

# 4. Feature dump per (ref, dist) pair using the existing harness:
ai/scripts/konvid_to_full_features.py \
    --corpus-root .workingdir2/bristol/bvi-cc \
    --manifest    ai/src/vmaf_train/data/manifests/bvi-cc.json \
    --backend     cuda \
    --out         ai/data/features/bvi-cc.parquet

# 5. Parquet lands at ai/data/features/bvi-cc.parquet (gitignored);
#    manifest stays in-repo, features stay out-of-repo.

Sizing: at ~125 ms/frame end-to-end on a single mid-range GPU and ~300 frames/clip, the 306 distorted clips are ~2.5 GPU-hours wall clock; doubled for both ref and dist features puts the soak at ~5–6 GPU-hours. Fits comfortably in an idle overnight slot.

Disk: assume ~250 GB on .workingdir2/bristol/bvi-cc (raw YUV plus the encoded distorted streams the dataset already ships pre-encoded). The dev box must have at least 400 GB free before download; otherwise stage onto an external NVMe and bind-mount the corpus root.

5. Risks¶

1. Storage blow-out. Pulling more than one BVI- set in sequence puts the box at TB scale fast. BVI-DVC alone is in the 700 GB – 1.2 TB band. Mitigation: ingest BVI-CC first, ship the manifest + parquet, then* decide whether to pull BVI-AOM next; never download two simultaneously.
2. MOS-scale mismatch. Bristol uses DMOS (difference MOS, higher = worse) for BVI-CC, BVI-HD, BVI-VFI; Netflix Public uses MOS (higher = better). Naïvely concatenating the two training sets without an inversion + re-scale step poisons the regressor. Mitigation: every parquet emits an explicit mos_convention column (netflix_mos, bristol_dmos, bristol_mos_inverted); the training loader normalises to a single 0–100 "higher is better" scale before fitting.
3. Licence misread on BVI-AOM CableLabs subset. Publishing a parquet derived from CC-BY-NC-ND 3.0 source is a "no-derivs" violation. Mitigation: the AOM ingest must consult a per-clip licence map and exclude restricted clips from any redistributable artifact.
4. Pre-encoded streams drift across BVI-CC versions. The distorted leg is a fixed encode of HM 16.18 / AV1 0.1.0 / VTM 4.01. If Bristol re-uploads with a newer encoder version, the parquet's bitrate / quality columns become inconsistent. Mitigation: pin the codec-version triplet inside the manifest and refuse to re-ingest a corpus root with mismatched names.
5. Form-gated download blocks reproducibility. Future contributors cannot reproduce the parquet without their own registration + ~2 day wait. Mitigation: ship the manifest in-repo (sequence list, expected SHA-256, MOS column) and let the parquet itself remain a personal-build artifact.
6. Parity-soak cost on cross-backend diff. Running /cross-backend-diff over 306 clips × 4 backends × full feature set is ~24 GPU-hours, not an interactive operation. Mitigation: a --bvi-cc-soak weekly CI run, not a per-PR gate.

6. Recommendation¶

Ingest BVI-CC first, behind a dedicated PR, sized for ~1 week of work:

1. Submit the registration form today (out-of-band, by the user).
2. While waiting, land the manifest scaffold and parquet schema change against an empty corpus root: bvi-cc.json manifest, mos_convention column added to the feature parquet, loader support for inverted-DMOS normalisation, ADR proposing the ingest, doc page under docs/ai/training-data.md.
3. When the OneDrive link arrives, fill the corpus root, run the feature dump, ship the parquet locally (gitignored), and run a confirmatory /cross-backend-diff over a 10-clip subset as a parity smoke test before any tiny-AI fitting.
4. Do not pull BVI-DVC or BVI-AOM in the same PR; that's a separate decision once we've measured fr_regressor_v2's lift from BVI-CC alone.

Companion ADR draft sits at docs/adr/0241-bristol-bvi-cc-ingest.md (Status: Draft, number unassigned) — it formalises the same recommendation and lists the alternatives we walked.

References¶

• Frontiers: https://www.frontiersin.org/journals/signal-processing/articles/10.3389/frsip.2022.874200/full
• BVI-CC site: https://fan-aaron-zhang.github.io/BVI-CC/
• BVI-DVC site: https://fan-aaron-zhang.github.io/BVI-DVC/
• BVI-AOM repo: https://github.com/fan-aaron-zhang/bvi-aom
• BVI-HFR site: https://fan-aaron-zhang.github.io/BVI-HFR/
• BVI-VFI repo: https://github.com/danier97/BVI-VFI-database
• BVI-RLV preprint: https://arxiv.org/abs/2407.03535
• Lab person page (gateway): https://research-information.bris.ac.uk/en/persons/david-r-bull/datasets/
• Companion: Research-0033 — Bristol VI-Lab NVC review (2026-04 preprint)
• Prior art on the fork: ADR-0042 (tiny-AI docs), ADR-0235 (codec-aware fr_regressor_v2), ADR-0019 (tiny-AI Netflix training)
• Memory: project_netflix_training_corpus_local.md (existing 37 GB Netflix Public corpus root)
• Source: req (user direction, 2026-05-02 — Bristol VI-Lab feasibility for tiny-AI training and parity soak)