SemanticMoments：通过第三矩特征实现无训练的运动相似性

Retrieving videos based on semantic motion is a fundamental yet unresolved challenge. Existing video representation methods overly depend on static appearance and scene context, neglecting motion dynamics, a bias inherited from their training data and objectives. Traditional motion-centric inputs, such as optical flow, lack the semantic grounding necessary for understanding high-level motion. To illustrate this inherent bias, the SimMotion benchmarks are introduced, combining controlled synthetic data with a new human-annotated real-world dataset. The results reveal that existing models perform poorly on these benchmarks, often failing to disentangle motion from appearance. To bridge this gap, SemanticMoments is proposed, a simple, training-free method that effectively captures motion information through third moment features, offering a new perspective on motion similarity retrieval.