MotionHalluc: Diagnosing Kinematic Hallucinations in Fine-Grained Motion Reasoning

The article introduces MotionHalluc, a benchmark designed to evaluate motion hallucinations in fine-grained motion reasoning across paired videos, consisting of 1540 questions derived from 553 video pairs. The benchmark assesses hallucinations in three dimensions: directional, attributional, and temporal, revealing that state-of-the-art multimodal models are highly prone to these issues. The authors propose a training-free method called Perceive-Parse-Verify (PPV) that enhances performance by integrating kinematic measurements during inference, achieving an average performance improvement of 10.6%, highlighting the importance of quantitative metrics in mitigating hallucinations in motion instruction generation.