The use of synthetic datasets to develop, prototype and qualify new computer vision algorithms is currently not widely accepted, though highly sought after by the industry. This is due to lack of knowledge on how the results acquired with such datasets will transfer to real live performance. Therefore, this paper introduces an approach to evaluate modelled synthetic datasets against their real counterparts. In a use case, the performance of common feature detectors is evaluated using the repeatability metric against real and synthetic datasets. Based on resulting performances; general usability, rendering techniques and modelling efforts for generation of synthetic datasets are discussed.
«The use of synthetic datasets to develop, prototype and qualify new computer vision algorithms is currently not widely accepted, though highly sought after by the industry. This is due to lack of knowledge on how the results acquired with such datasets will transfer to real live performance. Therefore, this paper introduces an approach to evaluate modelled synthetic datasets against their real counterparts. In a use case, the performance of common feature detectors is evaluated using the repeata...
»