Locating boundary is very important for Temporal Action Detection (TAD) and is a key factor affecting the performance of TAD. However, two factors lead to inaccurate boundary localization: the movement feature submergence and the existence of multi-scale actions. In this work, to address the submergence of movement feature, we design the Movement Enhance Module (MEM), in which the Movement Feature Extractor (MFE) and Multi-Relation Module (MRM) are used to highlight short-term and long-term movement information respectively. To address the characteristic of multi-scale actions, we propose a Scale Feature Pyramid Network (SFPN) to detect multi-scale actions and design a two-stage training strategy that makes each layer focus on a specific scale action. These tow modules are integrated as M3Net, and extensive experiments demonstrate its effectiveness. M3Net outperforms other representative TAD methods on ActivityNet-1.3 and THUMOS-14.