Automatically dataset generation using MOT analysis for deep metric learning

Abstract

Occlusion reduces the performance in Object Tracking. For keeping tracking when occlusion for a long period of time occurs, Metric Learning is usually used that measures similarity between images. On the other hand, it is difficult to train Metric Learning model because making training data takes high cost. In this paper, we propose the method for making training data automatically by MOT analysis and evaluate the model trained by this method.