MLT-Dedup: Efficient Large-Scale Online Video Deduplication via Multi-Level Representations and Spatial-Temporal Matching
Researchers have developed MLT-Dedup, a new framework designed to efficiently identify and remove near-duplicate videos from large online platforms. The system utilizes a novel Multi-Level Video Encoder to generate both detailed frame-level and broader clip-level embeddings, allowing for rapid candidate retrieval and precise matching. A key component, the Differential Feature-enhanced Similarity Module, accurately locates duplicated segments and provides evidence for deduplication decisions. Experiments show MLT-Dedup can reduce online repetition by 91% with 90% precision and significantly increases indexing capacity. AI
IMPACT Enhances efficiency and cost-effectiveness for platforms managing large video libraries by improving duplicate detection.