Explainable Forensics of Manipulated Segments in Untrimmed Long Videos

Researchers propose a new video forensics task focused on detecting, localizing, and explaining manipulated segments inside untrimmed long videos.

They release TASLE, a benchmark with 12,472 annotated videos designed to study AI-manipulated content hidden in mostly authentic footage.

The paper also introduces MSLoc, a coarse-to-fine baseline that first generates candidate proposals and then uses an MLLM-based refinement step for more precise localization and explanation.

The work fills a key gap in video forensics by adding both long-video temporal localization and explainability for human review.