Guided autoencoder for dimensionality reduction of pedestrian features

摘要

Autoencoder and other conventional dimensionality reduction algorithms have achieved great success in dimensionality reduction. In this paper, we present an improved autoencoder structure, which was applied it in the field of pedestrian feature dimensionality reduction. The novel method is also verified on Mnist dataset. High-dimensional deep pedestrian features outperform other descriptors while it is challenging for computing capability and memory in existing systems. The dimensionality reduction method we proposed takes advantages of autoencoder and principal component analysis to achieve high efficiency. A novel weight matrix initialization and an improved reconstruction of autoencoder are proposed. Furthermore, by fusing features labeled with the same pedestrian, the proposed structure minimizes the loss after dimensionality reduction. Experimental results demonstrate that our method outperforms traditional dimensionality reduction methods. In the experiment, the pedestrian features were generated by ResNet and Market-1501 data-set. Our method achieves up to 8.834% mAP increment compared to a principal component analysis, when 2048-dimension pedestrian features are reduced to 16-dimension features.