In order to achieve three-dimensional detection of artificial compound eye system and solve such problems as large field of view and complex imaging process, a global geometric calibration method based on two-cylinder and back-propagation (BP) neural network is proposed. In this method, the powerful mapping ability for arbitrary complex nonlinear relationship of BP neural network is used to take implicit calibration on two cylinders respectively, which set up the mapping relationship between the two-dimensional image coordinates and the cylindrical coordinates. In the three-dimensional measurement, the points on an image are mapped to two points on both the front and the back cylinders. Linking the two points we get a space line, and the intersections of multiple spatial lines are solved by the least square method. The coordinates in the three-dimensional target are then obtained. The experimental results show that the angle error of the artificial compound eye system is within 1 mrad at the field of view of 60°×30°, and the relative radial error is about 0.6%. Compared with traditional calibration methods based on pore model, this method does not need to consider compound eye imaging model and its parameters. Since the whole calibration process is conducted in the same world coordinates, the direction and location relationships between each eye need not to be accurately known. This method is fully applicable to geometric calibration of artificial compound eye systems, and it can meet the requirements of wide field and high precision calibration.