Data Augmentation and Fine Tuning of Convolutional Neural Network during Training for Person Re-Identification in Video Surveillance Systems

A new image set, augmentation method and fine in-learning adjustment of convolutional neural networks (CNN) are proposed to increase the accuracy of CNN-based person re-identification. Unlike other known sets, we have used many video frames from external and internal surveillance systems shot at all seasons of the year to make up our PolReID1077 set of person images. The PolReID1077-forming samples are subjected to the cyclic shift, chroma subsampling, and replacement of a fragment by a reduced copy of another sample to get a wider range of images. The learning set generating technique is used to train a CNN. The training is carried out in two stages. The first stage is pre-training using the augmented data. At the second stage the original images are used to carry out fine-tuning of CNN weight coefficients to reduce in-learning losses and increase re-identification efficiency. The approach doesn't allow the CNN to remember learning sets and decreases the chances of overfitting. Different augmentation methods, data sets and learning techniques are used in the experiments.