2 A New Sport Teams Logo Dataset for Detection Tasks
Andrey Kuznetsov, Samsung-PDMI Joint AI Center St. Petersburg, Andrey Savchenko, Samsung-PDMI Joint AI Center St. Petersburg, Higher School of Economics, Nizhny Novgorod
logo detection, sport logo, dataset design, VGG16, MobileNet, YOLO v3, SSD, deep learning
In this research we introduce a new labelled SportLogo dataset, that contains images of two kinds of sports: hockey (NHL) and basketball (NBA). This dataset presents several challenges typical for logo detection tasks. A huge number of occlusions and logo view changes during playing games lead to an ambiguity of a straightforward detection approach use. Another issue is logo style changes due to seasonal kits updates. In this paper we propose a two stage approach, in which, firstly, scene is recognition on an input photo using special-ly trained convolutional neural network. Second, conventional object detectors are applied only for sport scenes. Experimental study contains results of different combinations of backbone and detector convolutional neural networks. It was shown that MobileNet + YOLO v3 solution provide the best quality results on the designed dataset (mAP=0.74, Recall=0.87).