Objective Construction of Ground Truth Images

In the context of virtual and augmented reality, computer vision plays a pivotal role. To benchmark performance, evaluation of computer vision models, such as edge detection is essential. Traditionally this has relied on subjective analysis of the resultant images. Alternatively, models have been assessed against ground truth images. However, ground truth images are highly subjective, relying on human judging to determine the appropriate location of features. Literature complains about the lack of objective quantitative measures for model evaluation, yet no solution has been presented. Ground truth is the objective verification of properties of an image. In the context of this paper it is a data set that includes an accurate and complete representation of the edges. The subjective nature of creating ground truth images has meant that true image analysis model evaluation has been limited. Reducing the level of subjective decisions can improve the confidence level when measuring the performance of computer vision image analysis models. This work describes a new method to improve ground truth image confidence through an automated computer vision feature detection model voting system.