Research-0136: HDR/UGC Dataset License Audit (2026-05-15)¶

Status: Complete Date: 2026-05-15 Branch: research/hdr-ugc-dataset-license-audit-2026-05-15 Related ADR: ADR-0459 (panel-aware workstream) Companion task: Batch 16 of the 2026-05-15 Gap-Fill Plan

Summary¶

License, distribution-stability, and actionability assessment for the 13 HDR/UGC video quality dataset leads enumerated in Audit Slice C.7 (2026-05-15). Findings directly inform which datasets can be ingested into local training pipelines this week versus which require negotiation or infrastructure investment before they become useful.

Results: 6 datasets are ACTIONABLE-NOW, 5 are BLOCKED (access or license restrictions), 1 is ASPIRATIONAL-scale (needs infrastructure), and 1 (CHUG) is already active. HDRSDR-VQA introduced a new panel/display-aware workstream scoped in ADR-0459.

Dataset Summary Table¶

Per-Dataset Prose Assessment¶

1. Beyond8Bits¶

Beyond8Bits is the largest HDR-UGC corpus audited, with approximately 44,000 transcoded videos derived from 6,861 crowd-sourced source videos and over 1.5 million AMT quality ratings. The dataset is methodologically rigorous: ratings use SUREAL Maximum Likelihood Estimation aggregation, workers are screened for HDR-display ownership, and a golden-set calibration step filters low-quality annotators. All content is PQ/BT.2020 10-bit HEVC across a five-rung bitrate ladder (0.2–5 Mbps) at 360p, 720p, and 1080p — matching the CHUG transcoding philosophy closely.

The license situation requires care. Metadata is released under CC BY-SA 4.0, but the video payloads carry an additional "non-commercial research" restriction that prohibits redistribution outside the approved S3 mirror and deployment in commercial products without a separate license from UT Austin / YouTube. This NC clause is compatible with internal training but rules out shipping model weights that are demonstrably trained on this corpus without legal review, mirroring the existing BVI-DVC posture (weights ship locally only).

Frame rate coverage — 24/25/30/60 fps — does not include 120 fps, making it less affected by the HFR normalisation gap identified in CHUG's audit. It is the natural scale-up corpus to run after CHUG extraction completes.

2. CHUG (already active)¶

Already actively ingesting locally (chug_pipeline.pid running, 3,597 clips complete at audit time). Key open issues from Audit Slice H: missing --metadata-jsonl flag in the live invocation (content-split columns absent from the finished parquet), and the HFR normalisation gap (189 clips at 120 fps, ~525 at 60 fps, all processed without motion_fps_weight normalisation). Neither issue is blocked on external access; both are internal pipeline fixes tracked as open items. No further access work required.

3. YouTube SFV+HDR¶

The media.withyoutube.com/sfv-hdr landing page is JavaScript-rendered and returns only a title heading via WebFetch, consistent with the page being a research partner portal rather than a public download page. No paper preprint with a verified public arXiv ID was located. Based on secondary references in the HDR VQA literature, the dataset covers approximately 4,000 content items of which roughly 2,000 are native HDR; methodological details are not publicly documented. Access likely requires a YouTube Research partnership or signed data-sharing agreement. This dataset represents the highest potential scale in the audit cohort but is the most opaque regarding access terms. A direct contact with the YouTube Research team is the only viable path this week.

4. HDRSDR-VQA¶

This is the most strategically important dataset in the audit for the vmaf-tune toolchain specifically. The dataset covers 960 total video clips (558 with public open-source content; the remaining 145 subjects' VoD and live sports clips are unavailable due to copyright). The distinguishing feature is the six-display pairwise evaluation design: 145 participants rated HDR10 versus matched SDR versions on six distinct HDR televisions, producing scaled JOD scores. This introduces display-type variability data that no other dataset in the cohort provides, making it the unique resource for the panel-aware vmaf-tune training workstream scoped in ADR-0459. The license is permissive academic (no NDA, no royalty, citation required). Accessing the 31 open-sourced videos via Google Form is the immediate action. The 10 VoD clips and 10 live sports clips remain unavailable due to third-party copyright.

5. LIVE HDR Database¶

The LIVE HDR database provides 310 video clips from 31 source contents, evaluated by 66 participants (40 for the quality scoring task) under two ambient conditions. Its license is permissive academic (identical clause to HDRSDR-VQA: no royalty, no written agreement, cite the paper). The Google Form access gate is a minor friction; no NDA or institutional affiliation requirement is stated. The dataset is complementary to HDRSDR-VQA in that it studies ambient viewing condition effects (dark lab versus bright living room) rather than display panel variability. At 20,000+ opinion scores it is moderately sized and appropriate for validation of panel-aware or ambient-condition-aware quality models.

6. AVT-VQDB-UHD-2-HDR¶

This is the only 8K HDR corpus in the audit. The CC BY-NC 4.0 license permits non-commercial research use. The Google Form access gate implies active curation by TU Ilmenau. The principal limitation for immediate ingestion is infrastructure: 8K HDR decode requires significantly more compute than the 1080p CHUG clips, and the local NVIDIA GPU pipeline may not efficiently process 8K without additional buffering. The QoMEX 2024 paper should be reviewed before committing to download (124 GB S3 equivalent is the BVI-AOM reference point; this dataset is likely similar scale). The subjective methodology is not fully documented on the GitHub page; details are in the QoMEX paper. Not a week-1 priority but represents the only available 8K HDR quality corpus.

7. IPI-MobileHDRVQA¶

The simplest acquisition path of the 13: a CC BY 4.0 Zenodo record with direct download, no registration, 1.2 GB for the primary download (27.4 GB for this version, 197 GB total across all versions). The AV1 codec coverage is a gap relative to CHUG (HEVC-encoded), but AV1 represents an increasing fraction of real-world HDR delivery. The ACR-HR methodology with hidden reference is slightly richer than standard ACR, enabling DMOS computation. At 60 source videos it is small but immediately actionable without infrastructure investment. The mobile-oriented resolution mix (portrait mode content) adds a domain the CHUG pipeline does not cover. Download this week; run standard VMAF feature extraction against the AV1 bitstreams.

8. SHU-HDRQAD¶

This is an image quality (IQA) dataset, not a video quality dataset. At 1,409 HDR images across 147 scenes it has reasonable IQA scale, but the complete absence of an explicit license in the GitHub repository is a blocker: the Google Drive hosting provides no formal access controls but also provides no rights. Using unlicensed data for training model weights that ship in a commercial fork is legally risky. Additionally, the VQA training pipeline requires temporal content; a static image dataset has limited utility for training motion or compression-artifact quality models. The recommended action is to contact the maintainers requesting an explicit license grant — if they respond with CC BY or similar, the dataset could serve as an HDR IQA validation set. Skip for the near term.

9. BrightVQ¶

BrightVQ is architecturally very similar to Beyond8Bits and CHUG (same UT Austin + AMT + AWS S3 pipeline, CC BY-NC 4.0). At 73,794 ratings from 300 source videos it is substantially smaller than Beyond8Bits. The critical finding is that the full-dataset download link was marked "COMING SOON" at the time of audit (2026-05-15). This is not a license blocker — CC BY-NC 4.0 permits research use — but it is a practical access blocker. The dataset is from the same research group (Shreshth Saini, UT Austin) that published Beyond8Bits and CHUG, suggesting the data will become available through the same S3 mechanism. Monitor the GitHub repository; contact the author for early access if the timeline aligns with training needs.

10. BVI-AOM / BVI-CC¶

Despite the attractive technical specifications (10-bit, multi-resolution, 124 GB corpus, VMAF evaluation baseline), BVI-AOM's license explicitly restricts use to "developing future video coding standards" and "training or performance evaluation of test models in JVET and Alliance for Open Media." This restriction is incompatible with general vmaf-tune model training. The BVI-DVC corpus from the same Bristol VI Lab group is already ingested (PR #310, ADR-0310) under a research license; BVI-AOM adds no HDR content (it is confirmed SDR 10-bit) and carries a narrower license. Do not ingest BVI-AOM for the vmaf-tune pipeline.

11. SJTU HDR Video Sequences¶

SJTU provides 16 professional-grade HDR sequences captured with Sony F65/F55 cameras at 16-bit raw, stored as half-float 4:4:4 RGB OpenEXR after SMPTE ST.2084 tone mapping. The technical quality is exceptional — over 14 stops of dynamic range, BT.2020 primaries, S-Gamut/S-Log3. The access model (contact admin, download via Dropbox/file share) is manageable. However, two issues make this a low priority: (1) the CC BY-NC-ND 4.0 license's NoDerivatives clause explicitly prohibits format conversion, meaning the OpenEXR files cannot legally be decoded to YUV for VMAF feature extraction without constituting a "derivative work" under a strict reading; (2) there are no subjective quality ratings — the sequences serve as pristine HDR reference content for codec evaluation, not as a VQA training corpus. Contact SJTU to clarify the derivative-works scope before proceeding.

12. HDR-VDC¶

HDR-VDC (Cambridge gfxdisp lab, 2024) is a clean CC BY 4.0 corpus with 148 videos, 30 subjects, and pairwise JOD scores across AV1 compression and Lanczos upscaling distortions at three resolutions (720p, 1080p, 4K) on a high-quality LG G2 OLED display. The CC BY 4.0 license is the most permissive in the audit cohort — it permits redistribution and derivative works with attribution, making it suitable as a training corpus without the NC restrictions on weight shipping that Beyond8Bits and BrightVQ carry. The DOI-linked Cambridge Apollo repository provides a stable, archival download. The primary limitation is scale (30 subjects, 148 clips versus CHUG's 5,992). It is a strong validation set for the HDR-VDC quality dimension specifically (compression + upscaling artefacts on OLED). Download immediately.

13. AGH/NTIA/Dolby (CDVL)¶

The Consumer Digital Video Library (CDVL.org) requires free registration to access content. The public-facing pages state that content is "freely available for research and development purposes" once registered, but no specific license text is available without logging in. The AGH/NTIA/Dolby sequences are not listed on the public pages; their existence and HDR coverage must be verified post-registration. Dolby's involvement is a positive signal for HDR content but also introduces potential IP complexity. The CDVL is a well-established academic resource (long history of MPEG and HEVC evaluation content) and registration is low-friction. Register this week; search for the AGH/NTIA/Dolby collection; evaluate the license and HDR metadata before committing further.

Top 3 Priority Recommendations¶

Priority 1 — IPI-MobileHDRVQA (Zenodo CC BY 4.0) — Download this week¶

Why first: Zero friction. No registration, no sign-up, no NDA. CC BY 4.0 permits training use and derivative weight shipping without NC restrictions. At 60 source videos and 1.2–27 GB it fits on local storage, and AV1 content complements CHUG's HEVC coverage. Feature extraction can run using the existing CPU pipeline while the CHUG CUDA run continues.

Concrete action (this week): curl -L https://zenodo.org/records/11544387/files/<manifest> then run extract_k150k_features.py (NR mode) or wait for the CHUG chug_extract_features.py FR-pair adaptation to handle AV1 bitstreams. Open a BACKLOG entry: T-IPIHDRVQA-INGEST.

Priority 2 — HDRSDR-VQA (Google Form, permissive academic) — Apply today¶

Why second: Unique data. No other dataset in the cohort provides per-display-type pairwise HDR-vs-SDR quality scores. This data is the direct prerequisite for the panel-aware vmaf-tune workstream (ADR-0459). The Google Form gate is minutes of friction. The 31 open-sourced clips are available immediately after form submission; the 145-participant JOD score files likely accompany them.

Concrete action (this week): Submit Google Form at https://live.ece.utexas.edu/research/Bowen_SDRHDR/sdr-hdr-bowen.html; download the open-source 31-clip subset; extract VMAF features and JOD scores; design the panel-aware MOS-head schema (see ADR-0459 §Context).

Priority 3 — HDR-VDC (Cambridge Apollo, CC BY 4.0) — Download this week¶

Why third: Most permissive license in the cohort (CC BY 4.0, no NC restriction). The AV1 + Lanczos upscaling distortion types are complementary to CHUG's HEVC compression focus. The Cambridge Apollo DOI link (https://doi.org/10.17863/CAM.107964) provides archival stability. At 148 clips it is small but immediately usable as a validation holdout set or supplementary training shard.

Concrete action (this week): wget or curl from the Cambridge Apollo download page; add to BACKLOG as T-HDRVDC-INGEST. Extract VMAF features using existing CPU pipeline.

New Workstream Surfaced — Panel/Display-Aware vmaf-tune¶

Background¶

HDRSDR-VQA introduces a six-display pairwise evaluation design (OLED/QLED/LCD panels of varying peak luminance and color volume) that reveals substantial display-type-dependent perceptual quality differences for the same HDR10 + SDR video pair. This is not captured by any existing vmaf-tune model or recommendation workflow: the current vmaf-tune recommend path is display-agnostic — it assumes a generic high-end HDR display. In the field, HDR content is consumed on a range of panels where the same CRF setting can be perceptually optimal on one display and visibly degraded on another.

Proposed T-number¶

T-VMAFTUNE-PANEL-AWARE — Panel/display-aware vmaf-tune recommendation workstream.

Scope: Train a display-conditioned MOS/JOD head that accepts a display_type or peak_luminance_nits input feature alongside the standard VMAF feature vector. The head produces display-conditioned quality estimates enabling vmaf-tune recommend --display-profile {oled|qled|lcd|generic}. Data source: HDRSDR-VQA six-display JOD scores + HDRVDC OLED measurements as a second signal.

Estimated effort: T4 (medium: data ingestion 1–2 days, model head training 1 day, CLI integration 2 days, docs 1 day). Not blocked on any open PR.

ADR Reference¶

See ADR-0459 (Status: Proposed) for the full decision scaffold, alternatives considered, and consequences.