Sport Teams Logo Detection Based On Deep Local Features
Andrey V. Kuznetsov, A. Savchenko · 2019 International Multi-Conference on Engineering, Computer and Information Sciences (SIBIRCON) · 2019
Logo detection in an unconstrained environment is a challenging task. The problem lies in the small size of logos that are to be detected and different representations of logos due to the view angle, rotation angle and perspective distortions. We provide a part of a new labelled SportLogo dataset for research. In this work we also propose a solution to the task of sport logos detection for the whole range of MLB, NBA and HNL teams. Ambiguity appears when developing the detection algorithm because of a huge number of occlusions and logo view changes due to wrinkle uniform during hockey games. Another problem is logo style changes for some teams. First, we propose an approach based on local descriptors calculation (SIFT, SURF, ORB, BRISK, FREAK, AKAZE) and provide the results of conducted experiments. Second, we propose an approach based on deep local features (DELF), which are pre-trained on Google Landmarks dataset, and compare them with classic local features. Experimental research showed that BRISK features provide the highest quality values on the labelled NHL teams dataset.