GaitStrip: Gait Recognition via Effective Strip-Based Feature Representations and Multi-level Framework

Many gait recognition methods first partition the human into N-parts and then combine them to establish part-based feature representations. Their performance is often affected by partitioning strategies, which are empirically chosen in different datasets. However, we observe that strips as basic component of parts agnostic against strategies. Motivated this observation, present a strip-based multi-level network, named GaitStrip, extract comprehensive information at levels. To be specific, our high-level branch explores context sequences low-level one focuses on detailed posture changes. We introduce novel StriP-Based extractor (SPB) learn representations directly taking each strip body unit. Moreover, propose multi-branch structure, called Enhanced Convolution Module (ECM), gaits. ECM consists Spatial-Temporal (ST), Frame-Level (FL) SPB, has two obvious advantages: First, specific representation, can used improve robustness network. Specifically, ST aims spatial-temporal features sequences, while FL generate representation frame. Second, parameters reduced test introducing structural re-parameterization technique. Extensive experimental results demonstrate GaitStrip achieves state-of-the-art both normal walking complex conditions. The source code published https://github.com/M-Candy77/GaitStrip .