Academy & Industry Research Collaboration Center (AIRCC)

Volume 9, Number 3, March 2019

Geometric Deep Learned Feature Classification Based Camera Calibration


Cheolhyeong Park, Jisu Kim and Deokwoo Lee, Keimyung University, Republic of Korea


This paper chiefly focuses on calibration of depth camera system, particularly on stereo camera. Owing to complexity of parameter estimation of camera, i.e., it is an inverse problem the calibration is still challenging problem in computer vision. As similar to the previous method of the calibration, checkerboard is used in this work. However, corner detection is carried out by employing the concept of neural network. Since the corner detection of the previous work depends on the exterior environment such as ambient light, quality of the checkerboard itself, etc., learning of the geometric characteristics of the corners are conducted. The pro-posed method detects a region of checkboard from the captured images (a pair of images), and the corners are detected. Detection accuracy is increased by calculating the weights of the deep neural network. The procedure of the detection is de-tailed in this paper. The quantitative evaluation of the method is shown by calculating the re-projection error. Comparison is performed with the most popular method, Zhang’s calibration one. The experimental results not only validate the accuracy of the calibration, but also shows the efficiency of the calibration.


Calibration, Neural network, Deep learning, Re-projection error, Depth camera