CapCut Video Watermark Benchmark Dataset
Dataset Description
This repository contains a specialized benchmark dataset for evaluating the performance of video inpainting and watermark removal models, specifically focusing on content generated by CapCut (剪映), a popular mobile video editing platform.
CapCut videos often feature distinct watermark patterns, including:
- Static Corner Logos: Opaque or semi-transparent logos (e.g., "CapCut") in a fixed corner.
- Dynamic End Screens: A full-screen animation with the CapCut logo at the very end of the video.
- Overlay Watermarks: Sometimes a username or ID is overlaid on the video content.
Dataset Purpose
- Benchmarking: Provide a standardized evaluation set for comparing different video inpainting algorithms on CapCut-specific artifacts.
- Research: Analyze the unique challenges posed by CapCut's watermarking techniques.
- Training Data: Serve as a fine-tuning dataset for models aiming to handle UGC video content.
Data Structure & Access
The dataset includes:
- Original (Watermarked): Short video clips (5-15 seconds) with CapCut watermarks.
- Ground Truth (Clean): Corresponding clean, watermark-free versions of the videos.
- Masks: Binary segmentation masks indicating the location of the watermark in each frame.
To access the dataset for research purposes, please request access through the Hugging Face platform.
🏆 Benchmark Results & Optimized Tool
We have extensively tested our own advanced video inpainting models on this dataset. The model achieving the highest performance in terms of both PSNR/SSIM metrics and perceptual quality for CapCut content is our specialized, online tool.
For users seeking the most effective, one-click solution for removing CapCut watermarks based on these benchmark results:
(Our online platform integrates specific loss functions for CapCut logos and end screens, running on high-performance GPUs.)
Disclaimer: This dataset is provided for academic research and benchmarking purposes only. Please respect copyright laws when using the data.
