Improving object detection by exploiting semantic relations between objects
Tipus de documentProjecte Final de Màster Oficial
Condicions d'accésAccés obert
Object detection is a fundamental and challenging problem in computer vision. Detecting the objects visible in an image can give us a good understanding and description of the image. The extracted information can later be used to improve the results of other computer vision tasks like activity recognition, content-based image retrieval, scene recognition and more. As technology and internet connection are becoming more accessible, billions of people upload photos and videos every day. In order to make use of this enormous amount of data we need to be able to extract information from these images in a quick and yet reliable way. Convolutional neural networks (CNN) have made possible enormous progresses in object detection and classification in recent years and have already established themself as the state of the art approach for these problems. In this work, we try to improve object detection performances by employing a CNN approach able to exploit object co-occurrences in natural images. Typically, real world scenes often exhibit a coherent composition of object in terms of co-occurrence probability. For instance, in a restaurant we typically see dishes, bottles and glasses. We aim at using this type of knowledge as a cue for disambiguating object labels in a detection task.
En col·laboració amb la Universitat de Barcelona (UB) i la Universitat Rovira i Virgili (URV)