MFVSR
A large-scale, fine-grained evaluation benchmark for the video subtitle removal (VSR) task. It emphasizes an evaluation protocol based on spatial decoupling and information conditions, and covers multilingual and multi-style subtitles as well as diverse material such as authentic, forged, and low-resolution content.
Timeline (last 90 days)
MFVSR is a large-scale, multi-style benchmark for video subtitle removal oriented toward fine-grained and reliability evaluation.
The MFVSR benchmark contains 3,000 paired videos, covering three sources: Real-HQ (real high quality), Real-LQ (real low quality), and AI-Generated.
MFVSR also includes 953 in-the-wild real videos as an auxiliary set, used to analyze real-world coverage.
MFVSR covers 16 languages and characterizes subtitle differences along four dimensions: spatial layout, temporal behavior, visual style, and language pattern.
MFVSR's evaluation protocol distinguishes between two settings: blind (blind removal, no subtitle mask provided) and mask-guided (mask-guided removal).
MFVSR's evaluation protocol measures five categories of metrics: Restoration Fidelity, Preservation Fidelity, Temporal Fidelity, coverage, and failures.
The paper conducted experiments on six representative methods and found clear inconsistencies among local restoration quality, out-of-region preservation quality, temporal consistency, and full-set reliability.
The benchmark paper itself does not propose a new method, but redefines how the video subtitle removal task should be measured.
All Facts (8)
MFVSR is a large-scale, multi-style benchmark for video subtitle removal oriented toward fine-grained and reliability evaluation.
50%UnverifiedThe MFVSR benchmark contains 3,000 paired videos, covering three sources: Real-HQ (real high quality), Real-LQ (real low quality), and AI-Generated.
50%UnverifiedMFVSR also includes 953 in-the-wild real videos as an auxiliary set, used to analyze real-world coverage.
50%UnverifiedMFVSR covers 16 languages and characterizes subtitle differences along four dimensions: spatial layout, temporal behavior, visual style, and language pattern.
50%UnverifiedMFVSR's evaluation protocol distinguishes between two settings: blind (blind removal, no subtitle mask provided) and mask-guided (mask-guided removal).
50%UnverifiedMFVSR's evaluation protocol measures five categories of metrics: Restoration Fidelity, Preservation Fidelity, Temporal Fidelity, coverage, and failures.
50%UnverifiedThe paper conducted experiments on six representative methods and found clear inconsistencies among local restoration quality, out-of-region preservation quality, temporal consistency, and full-set reliability.
50%UnverifiedThe benchmark paper itself does not propose a new method, but redefines how the video subtitle removal task should be measured.
50%