Computer Vision News - January 2018

to their need. In addition to experience in the field, RSIP Vision offers clients its technical proficiency: for instance, solving registration problems using models and algorithms which allow to keep track of the interesting object in the following frames. Another registration technique is performed through feature extraction and tracking accompanied by machine learning algorithms : in this way, the software learns to adapt to any changes in the shape of the object due to the shift in gaze. The role of machine learning in this case is that of, given the registration of the following frame, learn the features seen in the new scene and use them in the next registration. Since a virtual object is located or moves in the scene following the coordinates system of the real object, tracking of this real object plays a significant role in AR applications . Tracking is an easy task when the tracked object has markers, but this case is often unfeasible or not practical. Tracking of markerless objects can be model-based or image-based: state- of-the-art methods utilize Machine Learning approaches for ongoing learning of the tracked object's shape, since it could change during the video. The main Machine Learning tracking approaches are two: a) CNN-based and b) Kernelized Correlation Filters - KCF-based. State-of-the-art CNN-based approach includes Recurrent Neural Networks for best absolute object location prediction and Reinforcement Learning for tracking instructions. Algorithms can also offer a solution in cases of partial occlusion of the tracked objects, boosting the reliability of the resulting view. Another key issue is dealing with changes in distance between the camera and the object: proportions need to be respected if we want to correctly identify the object in the new frame. That means that the algorithm needs to understand the dimensions of the objects in the scene, so that each can be rendered in the proper way. Computer Vision News “Proportions need to be respected if we want to correctly identify the object in the new frame” Project Computer Vision News 31 Image Processing Project “Software learns to adapt to any changes in the shape of the object due to the shift in gaze”