FiGVCL: Fine-Grained Benchmark and Method for Video Copy Localization

Area of Science:

• Computer Science
• Artificial Intelligence
• Multimedia Analysis

Background:

• Content-based video copy localization (VCL) is crucial for identifying edited copied video segments.
• Current VCL systems face challenges due to the high cost of annotation and the lack of fine-grained benchmarks.
• Robust VCL requires sophisticated video analysis to handle edited content.

Purpose of the Study:

• To address the limitations in VCL research by introducing a new dataset and evaluation metric.
• To facilitate the development of more effective and fine-grained video copy localization methods.
• To provide a benchmark for evaluating VCL systems on challenging, real-world scenarios.

Main Methods:

• Annotation of a new real-world dataset, FiGVCL, focusing on temporal correspondences in copied segments.
• Proposal of a novel fine-grained VCL benchmark metric utilizing temporal correspondences for enhanced discriminability.
• Development of a baseline model employing fine-grained local embeddings for precise copied segment localization.
• Implementation of an unsupervised training strategy for VCL.

Main Results:

• The proposed FiGVCL dataset enables evaluation of VCL methods in challenging scenarios.
• The new metric improves the discriminability of VCL systems.
• The baseline model achieves accurate copied segment localization.
• The unsupervised training strategy surpasses previous supervised VCL methods in performance.

Conclusions:

• The FiGVCL dataset, metric, and baseline models offer significant advancements for video copy localization research.
• Unsupervised learning presents a promising direction for developing effective VCL systems.
• The introduced resources will accelerate progress in detecting and localizing edited copied video segments.