Create README.md

README.md

@@ -0,0 +1,304 @@
+---
+pretty_name: LEHA-CVQAD
+tags:
+- video
+- computer-vision
+- compression
+- video-quality-assessment
+- subjective-quality
+- benchmark
+size_categories:
+- 1K<n<10K
+license: apache-2.0
+---
+
+# Dataset Card for LEHA-CVQAD
+
+## Dataset Summary
+
+LEHA-CVQAD is a large-scale dataset for **compressed video quality assessment**. It is designed for benchmarking and training both **full-reference (FR)** and **no-reference (NR)** video quality assessment methods on modern compression artifacts.
+
+The dataset combines:
+- **diverse source content**, including both professionally produced material and user-generated content,
+- **modern compression standards and codec presets**,
+- **pairwise preference annotations** converted into ranking scores,
+- **MOS / DMOS annotations** on a selected subset,
+- and a public **open split** plus a **hidden split** used for blind benchmark evaluation.
+
+This repository contains the **public open part** of the dataset. The hidden part is not released publicly and is used for benchmark evaluation to reduce overfitting.
+
+- Paper: https://arxiv.org/abs/2507.03990
+- Earlier dataset / methodology paper: https://arxiv.org/abs/2211.12109
+- Benchmark page: https://videoprocessing.ai/benchmarks/video-quality-metrics.html
+- Dataset / project page: https://videoprocessing.ai/datasets/cvqad.html
+
+## Supported Tasks and Leaderboards
+
+This dataset can be used for:
+
+1. **No-reference video quality assessment**
+   - Predict perceptual quality from a distorted video alone.
+
+2. **Full-reference video quality assessment**
+   - Predict perceptual quality from a distorted video and its pristine reference.
+
+3. **Pairwise ranking / preference learning**
+   - Learn relative quality ordering between compressed variants of the same source content.
+
+4. **Quality regression**
+   - Predict MOS, DMOS, Bradley-Terry scores, Elo scores, or a fused subjective score.
+
+5. **Codec / rate-distortion optimization research**
+   - Study how objective metrics align with human preference under bitrate constraints.
+
+Benchmark results for many IQA/VQA metrics are reported on the MSU benchmark website.
+
+## Languages
+
+The dataset is visual. Spoken language is not a primary annotation axis.
+
+## Dataset Structure
+
+### Data Instances
+
+A typical dataset instance represents one compressed video and its metadata.  
+Replace field names below with the exact keys used in your CSV / JSON metadata.
+
+```json
+{
+  "id": "leha_cvqad_000001",
+  "reference_id": "src_012",
+  "distorted_video": "distorted/codec_xxx/video_000001.mp4",
+  "reference_video": "references/src_012.y4m",
+  "split": "open",
+  "source_type": "raw_or_ugc",
+  "content_category": "sports",
+  "codec_family": "hevc",
+  "codec_name": "x265",
+  "preset": "medium",
+  "target_bitrate_kbps": 2000,
+  "width": 1920,
+  "height": 1080,
+  "fps": 30,
+  "bt_score": 0.73,
+  "elo_score": 1462.1,
+  "mos": 13.6,
+  "dmos": 5.2,
+  "fused_score": 0.69
+}
+````
+
+### Data Fields
+
+Use this section to describe your actual metadata schema. A typical release may contain:
+
+* `id`: unique sample identifier
+* `reference_id`: identifier of the source video
+* `distorted_video`: path or filename of the compressed video
+* `reference_video`: path or filename of the reference video for FR evaluation
+* `split`: dataset split, usually `open`
+* `source_type`: whether the source content is pristine / professional or UGC
+* `content_category`: coarse content label, if provided
+* `codec_family`: compression standard family (for example AVC, HEVC, VVC, AV1, VP9)
+* `codec_name`: concrete encoder / codec implementation
+* `preset`: encoding preset
+* `target_bitrate_kbps`: target bitrate used during encoding
+* `width`, `height`, `fps`: technical properties of the sample
+* `bt_score`: Bradley-Terry subjective ranking score
+* `elo_score`: Elo-based subjective ranking score
+* `mos`: mean opinion score
+* `dmos`: differential mean opinion score
+* `fused_score`: unified score derived from pairwise and rating experiments
+
+### Data Splits
+
+The full LEHA-CVQAD benchmark is divided into:
+
+* **Open split:** 1,963 videos
+* **Hidden split:** 4,277 videos
+
+The full benchmark contains 6,240 distorted videos in total.
+This Hugging Face repository releases the **open split only**.
+
+## Dataset Creation
+
+### Curation Rationale
+
+The dataset was created to support **generalized video quality assessment of compression artifacts**. Existing public datasets often have one or more of the following limitations:
+
+* limited codec diversity,
+* limited content diversity,
+* lack of authentic UGC artifacts,
+* small scale,
+* or subjective labels that are hard to compare across sources.
+
+LEHA-CVQAD was designed to address these issues by combining diverse source content, a wider range of codec presets, and a richer subjective annotation protocol.
+
+### Source Data
+
+#### Initial source video extraction
+
+The source-video extraction methodology follows the earlier CVQAD work. In that pipeline, high-bitrate open-source videos were collected from Vimeo, filtered by license and bitrate, converted to YUV 4:2:0, and sampled using spatial/temporal complexity clustering to obtain a representative and diverse set of reference videos.
+
+#### LEHA-CVQAD extension
+
+LEHA-CVQAD extends this approach by collecting a larger candidate pool of FullHD pristine videos from Vimeo and Xiph, adding more UGC content, and sampling a final set of **59** reference videos through complexity-aware clustering and manual selection for genre diversity.
+
+The final source set includes categories such as:
+
+* sports,
+* gaming,
+* nature,
+* interviews / television clips,
+* animation,
+* vlogs,
+* advertisements,
+* music videos,
+* water surfaces,
+* face close-ups,
+* and UGC.
+
+### Data Collection and Processing
+
+Each source video was compressed using a broad range of codecs and presets to cover modern compression artifacts. The full benchmark uses:
+
+* **186 codec / preset variants**
+* **3 target bitrates:** 1000, 2000, and 4000 kbps
+* multiple compression standards including AVC, HEVC, VVC, VP9, AV1, and others
+
+Not every source video was encoded with every available codec.
+
+The public benchmark design separates:
+
+* **open-source codec outputs** into the public split,
+* and **proprietary codec outputs** into the hidden split used for blind evaluation.
+
+### Annotations
+
+LEHA-CVQAD provides two kinds of subjective labels:
+
+1. **Pairwise rankings**
+
+   * Viewers compare two compressed videos derived from the same source and choose the better one or mark them as equivalent.
+   * These comparisons are converted into subjective ranking scores using **Bradley-Terry** and **Elo** models.
+
+2. **MOS / DMOS labels**
+
+   * A subset of videos is rated with **Absolute Category Rating (ACR)** on a **21-point scale**.
+   * Reference videos are included to support DMOS computation.
+
+The final dataset uses a fusion procedure to combine pairwise and rating information into a more globally consistent quality scale.
+
+### Subjective Study Protocol
+
+All subjective data were collected through a crowdsourcing platform in a browser-based full-screen setting. Videos were shown at FullHD resolution and pre-buffered before playback. Verification questions and participant filtering were used for quality control.
+
+For pairwise comparisons:
+
+* each task contained 12 videos presented as pairs,
+* two comparisons were verification questions,
+* at least 10 valid responses were collected for each pair.
+
+For MOS collection:
+
+* participants completed training before rating,
+* quality control filtered inconsistent or low-effort responses.
+
+### Statistics
+
+For the full LEHA-CVQAD benchmark:
+
+* **59** source videos
+* **6,240** distorted videos
+* **1,963** videos in the public open split
+* **4,277** videos in the hidden split
+* **186** codec / preset variants
+* approximately **1,797,310** valid pairwise responses
+* more than **15,000** unique participants in pairwise experiments
+* MOS responses collected from **1,496** participants
+
+## Dataset Use
+
+### Direct Use
+
+The dataset is suitable for:
+
+* training NR-VQA models,
+* training FR-VQA models,
+* metric benchmarking,
+* pairwise ranking models,
+* regression models for perceptual quality,
+* and research on codec optimization and rate-distortion alignment.
+
+### Out-of-Scope Use
+
+The dataset is not intended for:
+
+* general video understanding,
+* semantic recognition,
+* action recognition,
+* captioning,
+* or speech / language tasks.
+
+It should also not be treated as a universal proxy for all video distortions, since it is specifically oriented toward **compression-related quality assessment**.
+
+## Bias, Risks, and Limitations
+
+Several limitations should be considered:
+
+* The public release is only the **open split**. Results on this split alone may overestimate generalization compared with blind benchmark evaluation.
+* The dataset focuses on **compression artifacts** and is less suitable for unrelated distortions such as camera shake, defocus, or transmission artifacts unless they are already present in the source content.
+* Subjective studies were crowd-based rather than fully controlled laboratory studies.
+* Some content categories, codecs, or bitrate regions may be easier for current metrics than others.
+* Pairwise labels are naturally local to variants of the same source; the fused scale reduces but may not eliminate all cross-content comparability issues.
+
+## Data Preprocessing
+
+Typical preprocessing for research use may include:
+
+* decoding compressed videos to frames,
+* temporal subsampling,
+* patch or clip extraction,
+* normalization of subjective labels,
+* and pairing distorted videos with reference videos for FR evaluation.
+
+Researchers should report preprocessing choices clearly for reproducibility.
+
+## Evaluation
+
+Common evaluation protocols include:
+
+* Spearman rank correlation coefficient (SRCC)
+* Pearson linear correlation coefficient (PLCC)
+* Kendall rank correlation coefficient (KRCC)
+* pairwise ranking accuracy
+* codec-wise or bitrate-wise subgroup evaluation
+
+When reproducing benchmark results, users should ensure that they do not train on hidden benchmark content.
+
+## Citation
+
+If you use this dataset, please cite the LEHA-CVQAD paper and the earlier CVQAD methodology paper:
+
+```bibtex
+@article{gushchin2025leha,
+  title={LEHA-CVQAD: Dataset To Enable Generalized Video Quality Assessment of Compression Artifacts},
+  author={Gushchin, Alexander and Smirnov, Maksim and Antsiferova, Anastasia and others},
+  journal={arXiv preprint arXiv:2507.03990},
+  year={2025}
+}
+```
+
+```bibtex
+@inproceedings{antsiferova2022video,
+  title={Video compression dataset and benchmark of learning-based video-quality metrics},
+  author={Antsiferova, Anastasia and Lavrushkin, Sergey and Smirnov, Maksim and Gushchin, Alexander and Vatolin, Dmitriy and Kulikov, Dmitriy},
+  booktitle={NeurIPS 2022 Datasets and Benchmarks Track},
+  year={2022}
+}
+```
+
+## Additional Information
+
+This repository corresponds to the public dataset release.
+For blind evaluation on the hidden split and benchmark results for existing metrics, see the MSU benchmark pages linked above.