CrossSiam: k-Fold Cross Representation Learning

Abstract

One of the most important tasks for multi-agents such as drones is to automatically make decisions based on images captured by on-board cameras. These agents must be highly accurate and reliable. For this purpose, we applied k-fold cross validation to the task of classifying images using deep learning, which is a method that compares and evaluates models appropriately model of a given problem; this technique is easy to understand and easy to implement, and it produces results in lower bias estimates. However, k-fold cross validation reduces the amount of data per neural network, which reduces the accuracy. In order to address this problem, we propose CrossSiam. CrossSiam is a one of the representation learning methods to train feature encoders to mimic the embedding space of the validation data of each neural network. We show that the proposed method has a higher classification accuracy than the ParaSiam (baseline). This approach can be very important in the field where reliability is required, such as automated vehicles and drones in disaster situations.