視覚マー力を用いた物体検出のための偏りのない学習データ自動収集システム

－変種変量生産の自動化に向けて－

抄録

To accurately recognize various products in automated factories, deep learning-based vision systems are widely employed. Such systems require a significant amount of training data with manual annotations. Unfortunately, the manual annotation is very time consuming and error-prone. In order to solve this problem, we have proposed a method for automatic annotation using a single visual marker. With our method, the annotation time was reduced. On the other hand, it is necessary to shorten the image capturing time and collect a wide variety of training data.

In this paper, we propose an automatic system for collecting a wide variety of training data in a short time. The proposed system generates a training dataset with no bias in the data quantity of each position and orientation of target objects in a shorter time than manual method. The experiments verified the effectiveness of the proposed system by comparison with the manual method in both the time to collect training data and the accuracy of the vision system.