Impact of Camera Production Tolerances on Computer Vision
A company must be confident that the optical system in their product design can deliver performance requirements. In many cases, the production of faulty cameras is a difficult problem where, for example out of one million cameras produced in one day, 1% of this is discarded or taken apart and reproduced because they fail postproduction tests. This 1% error rate which accounts for 10,000 cameras is a significant number, and this is generally higher in industrial settings due to the increasing volume of production. Therefore, it is important to investigate how well cameras generalize on computer vision algorithms.
In these circumstances, photometric effects and optical characteristics of the lens, and in particular, the range of possible focus scores must be considered. Optical Performance parameters include 50% Modulation Transfer Function (MTF50) and Spatial Frequency Response in general which can be used interchangeably with MTF.
The autonomous industry has advanced considerably in the application of state-of-the-art computer vision systems with various optical designs in the interest of enhancing road safety. When using camera and sensor systems on vehicles there is a particular focus on a wide field-of-view to capture the entire surroundings of the vehicle. However, one crucial challenge in designing surround-view cameras is the strong radial aberration of the camera where the current algorithms prove difficult to train on this data. Current methods include the warping of video data where the distortion or spatial properties of the camera are removed for processing. This approach proves inefficient and crucial data is lost. Hence, it is interesting to look at variations in camera production tolerances to bridge the gap between computer vision algorithms and image distortion.