Beyond Local Cycles: Long-Horizon Repetitive Action Counting via Multi-Scale Temporal Representation Learning

Repetitive action counting aims to estimate the number of action repetitions in videos and is an important component of fine-grained video understanding. However, existing methods are mainly designed for relatively short video clips and often struggle with long-horizon videos, where repetitions exhibit large temporal variations, complex background changes, and long-range dependencies. In this work, we study repetitive action counting from the perspective of long-range temporal representation learning. We propose a unified framework that captures both local repetitive patterns and global temporal context, enabling the model to maintain reliable counting representations over extended video sequences. By modeling temporal information at multiple scales, our approach is designed to distinguish repetitive actions from irrelevant motions while remaining robust to variations in action speed and duration. Extensive experiments on repetitive action counting benchmarks demonstrate the effectiveness of the proposed framework for both standard and long-horizon counting scenarios. Our work provides a general perspective on scalable temporal modeling for repetition understanding in long videos.