Lightweight Object Detection-Tracking using Deep Feature Distillation

Object detection and tracking are critical and fundamental problems in machine vision task. In this paper, an object detection and tracking method is proposed based on deep feature distillation. Particularly, an adaptive unsupervised Teacher-Student unified framework is developed. The Teacher module is performed by an expandable generative adversarial network mixture model. And knowledge discrepancy ranking (KDR) is designed to optimize Teacher resource allocation with the historical underlying knowledge. The Student module is developed based on a lightweight probabilistic generative model. And an unsupervised learning scheme is presented based on Gumbel-Soft sampling optimization to train jointly. A series of experiments are performed on authoritative dataset, demonstrating that the proposed method outperforms the state-of-the-art comparison methods.