Latent training for convolutional neural networks

Abstract

Pedestrian detection and recognition has become the basic research in various social fields. Convolutional neural networks have excellent learning ability and can recognize various patterns with robustness to some extent distortions and transformations. Yet, they need much more intermediate hidden units and cannot learning from unlabeled samples. In this paper, we purpose a latent training model based on the convolutional neural network. The purposed model adopts part detectors to reduce the scale of the intermediate layer. It also follows a latent training method to determine the labels of unlabeled negative parts. Last, a two-stage learning scheme is purposed to overlay the size of the network step by step. Experimental results on the public static pedestrian detection dataset, INRIA Person Dataset [1], show that our model achieves 98% of the detection accuracy and 95% of the average precision.