Subjective and Objective Quality Assessment of HDR-UGC Videos CVPR 2026
Shreshth Saini¹, Bowen Chen¹, Neil Birkbeck², Yilin Wang², Balu Adsumilli², Alan C. Bovik¹,³ ¹ Laboratory for Image and Video Engineering (LIVE), UT Austin · ² Google / YouTube · ³ University of Colorado Boulder
Beyond8Bits is, to our knowledge, the largest crowdsourced HDR User-Generated Content (UGC) video quality dataset to date, paired with HDR-Q, the first multimodal large language mode[...]
| Paper-reported (full) | Publish-ready release | |
|---|---|---|
| HDR source videos | 6,861 (2,253 Crowd + 4,608 Vimeo) | 5,917 (2,153 Crowd + 3,764 Vimeo) |
| Transcoded videos | ~44,276 | 41,419 (12,918 Crowd + 22,584 Vimeo transcodes + 5,917 references) |
| Crowd ratings | >1.5 M on Amazon Mechanical Turk | ~1.46 M |
| Avg. ratings per video | ~35 | ~35 |
| Resolutions | 360p / 720p / 1080p (+ source) | same |
| Bitrate ladder | 0.2 – 5 Mbps | 0.2 / 0.5 / 1 / 2 / 3 Mbps |
| Rating instrument | Continuous 0–100 Likert, ITU-R BT.500-14 | same |
| Split | 70 / 20 / 10 (train / val / test), by source | 70 / 10 / 20 (28,987 / 4,151 / 8,281) |
| MOS aggregation | SUREAL MLE; median inter-subject SRCC 0.90 | same |
Beyond8Bits/
├── index.html # Project page (hosted via GitHub Pages)
├── images_data.json # Carousel: frame paths + predicted / MOS scores
├── static/
│ ├── css/ # bulma + project stylesheet
│ ├── js/ # fontawesome
│ ├── images/ # overview figure, distribution plots, AMT UI
│ └── pdfs/ # paper & supplementary (drop-in once camera-ready)
├── images/
│ └── sample_frames/ # 99 sample HDR frames used by the page carousel
└── data/
├── Beyond8Bits_publish.csv # Full release: 41,419 transcoded clips × 12 columns
├── Beyond8Bits_publish.txt # Matching list of 41,419 hashed video IDs
├── Beyond8Bits_publish_crowd.csv # CHUG-compatible crowd-only subset (5,992 rows)
└── Beyond8Bits_publish_crowd.txt # Matching crowd-only ID list
Per-rating raw CSVs and the full archive live on UT-Box: https://utexas.box.com/s/pvz8zpmpogvpy62pqpar2e54c6ovyd5z
data/Beyond8Bits_publish.csv (41,419 rows × 12 columns):
| Column | Description |
|---|---|
video_id |
Hashed video ID (primary key, used for S3 download) |
mos |
MOS after SUREAL bias-corrected aggregation |
sos |
Standard-of-scores (SUREAL dispersion) |
type |
Source partition, Crowd or Vimeo |
ref |
1 if source / reference video; 0 if transcoded |
resolution |
Target resolution, 360p / 720p / 1080p / ref (source) |
bitrate |
Target bitrate, 0.2M / 0.5M / 1M / 2M / 3M (or ref) |
orientation |
Portrait or Landscape |
framerate |
Native playback framerate (fps) |
split |
train / validation / test (70 / 10 / 20 by source identity) |
height |
Native frame height (px) |
width |
Native frame width (px) |
Per-rating raw CSVs are available via the UT-Box mirror linked above. See the paper Appendix E for the rating protocol.
Videos are hosted on an S3 mirror (s3://ugchdrmturk/videos/). You can use either the AWS CLI or stream directly in a browser.
# clone
git clone https://github.com/shreshthsaini/Beyond8Bits.git
cd Beyond8Bits
# single video
aws s3 cp s3://ugchdrmturk/videos/VIDEO_ID.mp4 ./Beyond8Bits_Videos/
# bulk: all 41,419 IDs in the published manifest
cat data/Beyond8Bits_publish.txt | while read video; do
aws s3 cp s3://ugchdrmturk/videos/${video}.mp4 ./Beyond8Bits_Videos/
doneTo stream a single video directly in the browser, replace VIDEO_ID in:
https://ugchdrmturk.s3.us-east-2.amazonaws.com/videos/VIDEO_ID.mp4
Example: https://ugchdrmturk.s3.us-east-2.amazonaws.com/videos/b882a56f87be722f24206e007c31ee4c.mp4
import pandas as pd
df = pd.read_csv("data/Beyond8Bits_publish.csv")
print(df.columns.tolist())
# ['video_id', 'mos', 'sos', 'type', 'ref', 'resolution',
# 'bitrate', 'orientation', 'framerate', 'split',
# 'height', 'width']
print(df["mos"].describe()) # MOS distribution
print(df["type"].value_counts()) # Crowd (15,071) vs Vimeo (26,348)
print(df["split"].value_counts()) # train / validation / test
print(df["resolution"].value_counts()) # ref / 360p / 720p / 1080pPaired with the dataset, we release HDR-Q, the first MLLM designed for HDR-UGC VQA.
- HDR-aware vision encoder: a SigLIP-2 encoder adapted with HDR–SDR dual-domain contrastive supervision (captions generated by Qwen2.5-VL-72B). Frames stay at native 10-bit PQ / BT.2020;[...]
- HAPO (HDR-Aware Policy Optimization): extends GRPO with three HDR-specific terms:
- HDR–SDR contrastive KL: prevents modality neglect by contrasting rollouts with and without HDR tokens.
- Dual-entropy regularization: per-pathway entropy penalties to avoid reward-hacking via entropy inflation.
- High-Entropy Weighting (HEW): rescales group-normalized advantage with per-token entropy, focusing gradients on informative reasoning tokens.
- Rewards: Gaussian-weighted MOS regression reward
R_sc+ format rewardR_fmt+ group-level self-rewardR_self. - Backbone + compute: Ovis2.5 base with rank-4 LoRA adapters; 8 uniformly-sampled frames per clip; trained on 4× NVIDIA H200 GPUs; two-stage RL (modality alignment → full RFT).
- Metadata, CSV and TXT manifests, and the website code: CC BY 4.0, as stated in the paper. Use, share, and build on them with attribution.
- Video payloads keep their original terms: CC-licensed Vimeo content, and crowd-contributed videos released under a non-exclusive research redistribution agreement (non-commercial research use). For commercial use of the videos, reach out.
@InProceedings{Saini_2026_CVPR,
author = {Saini, Shreshth and Chen, Bowen and Wang, Yilin and Birkbeck, Neil and Adsumilli, Balu and Bovik, Alan C.},
title = {Seeing Beyond 8bits: Subjective and Objective Quality Assessment of HDR-UGC Videos},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2026},
pages = {15538-15549}
}Please also consider citing the predecessor sub-studies Beyond8Bits extends:
@INPROCEEDINGS{Saini_2025_ICIP_CHUG,
author = {Saini, Shreshth and Bovik, Alan C. and Birkbeck, Neil and Wang, Yilin and Adsumilli, Balu},
booktitle = {2025 IEEE International Conference on Image Processing (ICIP)},
title = {CHUG: Crowdsourced User-Generated HDR Video Quality Dataset},
year = {2025},
pages = {2504-2509},
doi = {10.1109/ICIP55913.2025.11084488}
}
@InProceedings{Saini_2026_WACV_BrightRate,
author = {Saini, Shreshth and Chen, Bowen and Wang, Yilin and Birkbeck, Neil and Adsumilli, Balu and Bovik, Alan C.},
title = {BrightRate: Quality Assessment for User-Generated HDR Videos},
booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
month = {March},
year = {2026},
pages = {1522-1532}
}Main PoC: Shreshth Saini, saini.2 at utexas.edu · website · Google Scholar
Future Data Maintainer: Pragyadipta Adhya, pragyadipta.adhya at colorado.edu · website · Google Scholar