FAST-MDL: Fast Adaptive Supervised Training of multi-layered deep learning models for consistent object tracking and classification

In this paper, we propose a Fast Adaptive Supervised Training Algorithm, called FAST-MDL, for dynamically updating the parameters of a multi-layered deep learning structure in order to fit the current environmental conditions. The method provides promising results in consistent long term object labeling and detection under abruptly changing visual conditions, severe illumination changes and occlusions (either full or partial). The retraining algorithm trusts as much as possible the current conditions (discriminative constraints), while simultaneously providing a minimal modification of the already obtained knowledge of the network (generative constraints). Experimental results in real-life video sequences demonstrate the efficiency of the proposed method compared to existing techniques.