The application of Deep Learning Neural Networks to Object Detection grants the localization of object instances from several classes within images. The inference of the locations of the items is not always precise and the positional error of bounding boxes can result being in the pixel range. This research aims at analysing the object localization uncertainty of Deep Convolutional Neural Networks (DCNNs), by considering MVTec HALCON and open-source Neural Networks. The metrological outcomes are presented. Since matching algorithms are still used to identify shapes, contours and edges, a comparison with them sounds like an obligation. The results of both Deep Learning and Shape-Based Matching (SBM) algorithms are compared with a sub-pixel precise reference labelling. The labelling consists of specific bounding boxes around the objects of each image of the dataset. It was obtained manually performing rectangular boxes on a single image and propagating them to the entire dataset according to the camera pose, precisely determined by keeping the target fixed and moving the camera with a measuring machine. The intent of this work is to establish the best detection methodology with respect to different testing conditions, considering also the effect of possible Gaussian noise and imprecise labelling.
|