Deep Unsupervised Learning for Simultaneous Visual Odometryand Depth Estimation: A Novel Approach

Advancing Simultaneous Visual Odometry and Depth Estimation withDeep Unsupervised Learning: ICIP 2019 Conference Paper

: As known that the Co-training algorithm in semi-supervised learning is a classical algorithm in semi-supervised learning algorithm, but it needs sufficient redundant view of data set in the process of learning, but this requirement usually can not be met in actual application. There are some improved algorithms, one of which is Tri-training algorithm and it uses three classifiers. It is an important algorithm used in Semi-supervised Learning. Although it no longer requires the attribute set of data, its time complexity is high. In the paper, it improves the Co-training algorithm thorugh using two classifiers. The difference is that one classifier only trains its own labeled data, one classifier trains its own labeled data and some unmarked data with good confidence, and one classifier is supervised learning and the other is semi-supervised learning. Then, the two classifiers are used to predict all unlabeled data, and the unlabeled data with high confidence and prediction tags are added to the training set of classifier 2. Experiments show that the accuracy and running time of the proposed algorithm in image classification can be significantly improved.