Exploring the Simplification Limit of Deep Network Features With Subway Positioning Task

Abstract

This paper addresses vision-based subway positioning, a significant yet challenging task due to the low-lighting and sparse-texture conditions in tunnels. Traditional features struggle with temporal correspondence. While deep network features are effective, their computational and storage demands make them unsuitable for on-board systems. We propose a simple-structured feature extractor, trained via a student-teacher distillation framework to inherit the powerful pattern mining and abstraction capabilities of deep networks. Our goal is to simplify deep network features for fixed-route applications like subway positioning, developing an on-board efficient feature extractor for practical applications. Specifically, we design a single-layer convolution operator as our student network. Through discriminability augmented distillation, we compress the feature extraction capabilities of the state-of-the-art SiLK into this compact model, achieving an optimal balance between descriptive power and computational efficiency. Our method achieves a model size of 2 KB and a processing speed of 1453 FPS, while maintaining high homography estimation accuracy comparable to those of deep network features. Extensive experiments on the vision-based subway positioning dataset show our method offers superior speed and deployability without losing accuracy.