🎬 MFVSR

A Multi-Style and Fine-Grained Benchmark for Video Subtitle Removal

📄 Paper  ·  💻 GitHub

MFVSR provides 3,000 paired videos with subtitled inputs, clean ground truth, glyph masks, bounding-box annotations, BBox mask videos, and PNG mask sequences.

📦 Download

25 TAR archives · 109.55 GiB

pip install -U huggingface_hub
hf download siiian/MFVSR --repo-type dataset --local-dir MFVSR-download
cd MFVSR-download
sha256sum -c SHA256SUMS
mkdir -p ../MFVSR
for archive in archives/*.tar; do
  tar -xf "$archive" -C ../MFVSR
done
cp -r manifests ../MFVSR/

Keep the extracted paths unchanged. manifests/benchmark_manifest.json maps all assets; sample_archive.csv maps samples to archives.

Evaluation lists: manifests/lite3000.txt contains all 3,000 samples; manifests/audit300.txt defines the fixed 300-sample subset.

Download a subset (optional)

For a quick check or preliminary evaluation, you can download a source subset instead of the full dataset. AI-Generated contains 1,000 paired samples in 3 TAR archives (10.64 GiB).

Example: download and prepare AI-Generated

pip install -U huggingface_hub
hf download siiian/MFVSR --repo-type dataset --local-dir MFVSR-AI-download \
  --include "archives/AI-Generated-*.tar" --include "manifests/*" \
  --include "sample_archive.csv" --include "archive_index.csv" \
  --include "SHA256SUMS" --include "LICENSE" --include "DATA_SOURCES.md" \
  --include "UPSTREAM_LICENSES/*"
cd MFVSR-AI-download
sha256sum --ignore-missing -c SHA256SUMS
mkdir -p ../MFVSR-AI
for archive in archives/AI-Generated-*.tar; do
  tar -xf "$archive" -C ../MFVSR-AI
done
cp -r manifests ../MFVSR-AI/

python - <<'PY'
import csv
from pathlib import Path

with Path("manifests/sample_metadata.csv").open(encoding="utf-8", newline="") as f:
    ids = [row["video_id"] for row in csv.DictReader(f) if row["source"] == "AI-Generated"]
target = Path("../MFVSR-AI/manifests/ai_generated.txt")
target.write_text("\n".join(ids) + "\n", encoding="utf-8")
print(f"Saved {len(ids)} sample IDs to {target}")
PY

Set the evaluation toolkit's data root to MFVSR-AI and use --ids-file /path/to/MFVSR-AI/manifests/ai_generated.txt to evaluate only these samples.

For Real-HQ or Real-LQ, replace the archive pattern with archives/Real-HQ-*.tar or archives/Real-LQ-*.tar, and change the source filter and output names accordingly. For a custom selection, filter manifests/sample_metadata.csv by source, resolution, language, or subtitle style, then look up the required archives in sample_archive.csv. Downloads are at TAR-archive level; even a small selection may require multiple archives.

Report custom-subset results with the sample count and selection rule, and share the sample-ID list. To reproduce the paper's results, use the corresponding fixed evaluation lists and protocol.

🌍 Sources

Subset	Videos	Original video sources
Real-HQ	1,000	OpenVidHD-0.4M, Pexels
Real-LQ	1,000	Panda-70M, Pexels
AI-Generated	1,000	VBench-2.0 sampled videos, Rapidata Sora

Subtitle text comes from FLORES-200 and fonts from Google Fonts. See Data sources and terms for details. The paper's 953 auxiliary in-the-wild videos are excluded from this release.

⚖️ License

See MFVSR Dataset Terms. MFVSR-authored contributions covered there use CC BY 4.0; video content and subtitle text retain upstream terms, including Panda-70M's non-commercial research restrictions. Source-specific terms and open licensing details apply.