Gait Recognition with Mask-based Regularization

Most gait recognition methods exploit spatial-temporal representations fromstatic appearances and dynamic walking patterns. However, we observe that manypart-based methods neglect representations at boundaries. In addition, thephenomenon of overfitting on training data is relatively common in gaitrecognition, which is perhaps due to insufficient data and low-informative gaitsilhouettes. Motivated by these observations, we propose a novel mask-basedregularization method named ReverseMask. By injecting perturbation on thefeature map, the proposed regularization method helps convolutionalarchitecture learn the discriminative representations and enhancesgeneralization. Also, we design an Inception-like ReverseMask Block, which hasthree branches composed of a global branch, a feature dropping branch, and afeature scaling branch. Precisely, the dropping branch can extract fine-grainedrepresentations when partial activations are zero-outed. Meanwhile, the scalingbranch randomly scales the feature map, keeping structural information ofactivations and preventing overfitting. The plug-and-play Inception-likeReverseMask block is simple and effective to generalize networks, and it alsoimproves the performance of many state-of-the-art methods. Extensiveexperiments demonstrate that the ReverseMask regularization help baselineachieves higher accuracy and better generalization. Moreover, the baseline withInception-like Block significantly outperforms state-of-the-art methods on thetwo most popular datasets, CASIA-B and OUMVLP. The source code will bereleased.