draft.pdf (2.57 MB)
Download file

A benchmark image dataset for industrial tools

Download (2.57 MB)
journal contribution
posted on 19.03.2020, 12:15 by Cai Luo, Leijian Yu, Erfu Yang, Huiyu Zhou, Peng Ren
Robots and Artificial Intelligence (AI) play an increasingly important role in manufacture. One of the tasks is to identify tools in the scene so that the tools can be applied to different assembly purposes. In the AI community, many datasets have been generated and deployed to train robots to recognize individual items, however, these datasets are scene-specific and lack generic background. In this paper, we report our dataset contains photos of 8 objects types that would be easily recognized by qualified workers. This is achieved by gathering images of common tools in a typical factory. The ground truth categories of our dataset are manually labeled by experienced workers, which would be worthy evaluation tools for the intelligence industrial systems. The equipment used and the image collection process are discussed, along with the data format. The mean average precisions range from 64.37% to 78.20%, which bring the possibility for future improvement. The dataset is ideal to evaluate and benchmark view-point variant, vision-based control algorithm for industry robots. It is now public available from https://github.com/tools-dataset/Industrial-Tools-Detection-Dataset.

History

Citation

Pattern Recognition Letters, Volume 125, 2019, P 341-348

Author affiliation

Department of Informatics

Version

AM (Accepted Manuscript)

Published in

PATTERN RECOGNITION LETTERS

Volume

125

Pagination

341 - 348

Publisher

ELSEVIER

issn

0167-8655

eissn

1872-7344

Acceptance date

16/05/2019

Copyright date

2019

Available date

17/05/2019

Publisher version

https://www.sciencedirect.com/science/article/pii/S0167865519301606

Language

English