From Synthetic Images Towards Detectors of Real Objects: A Case Study on Road Sign Detection

摘要

Large amounts of suitably marked images are required to feed a learning algorithm when building a detector. The process of collecting images and then marking locations of target objects in them is arduous. One could potentially speed up this process if real images for a given application were replaced by images generated synthetically, and coordinates of targets were simply imposed, rather then discovered manually. Despite the appeal of such an automatization, questions arise regarding the usefulness of systems built this way in real operating conditions. In particular, the obvious violation of i.i.d. principle might result in higher error rates at the testing stage. In this paper we provide an experimental study of the above approach, taking up road sign detection as an example. We generate synthetic training scenes by laying road sign icons randomly over a set of backgrounds with additional perturbations (rotations, brightness changes, blurring, sharpening, noise). Ensemble learning is then carried out using a RealBoost algorithm with shallow decision trees. Haar-like features or Fourier moments constitute the direct input information extracted from images. In both cases we support the computations with suitable integral images. Accuracy of resulting detectors is finally tested on real images.