Gait recognition plays a crucial role in long-distance, unconstrained identity verification. Silhouette-based methods are widely recognized for their accuracy but often utilize single-stage temporal compression techniques, such as global pooling, to condense sequence temporal information into fixed-length features. However, this approach restricts their ability to fully capture the dynamic temporal characteristics inherent in gait sequences. To address this, we introduce SF-Gait, a novel framework employing two-stage compression-leveraging a 3D CNN for micro-motion and a 2D CNN for gait cycle features. It then anchors dense short-term micro-motion features to representative 2D contour features, forming sparse sequences as the initial stage of temporal compression. A subsequent global temporal pooling method, akin to other approaches, serves as the second stage of compression, thereby yielding more representative temporal features. Our Spatio-Temporal Downsampling (ST-D) and Dual-stream Fusion Encoder (DFE) enhance gait modeling capabilities, achieving state-of-the-art performance on CASIA-B, OU-MVLP, and Gait3D datasets.