GeLoGo: Detecting TV Logos from Web-Scale Videos

An efficient way to identify the source of a web video is to detect some specified logos (e.g. a TV logo or a brand). Ln this paper, we propose a logo detection system "GeLoGo" for web-scale videos via four major modules. The keyframe module first extracts several frames with shot segment detection. In this manner, the data can be dramatically reduced. After that, the proposal module localizes candidate logos from the extracted frames by an ensemble of some pretrained logo detectors. Then, the spatial verification module takes the logo proposals as input and identifies them as real logos or not by performing the classification task within a ResNet network. Finally, the temporal verification module further identifies the detection results by checking the temporal consistency of logo locations. Experimental results on a collected large-scale TV logo video dataset show that the proposed approach can achieve a detection accuracy of 99.1% with a speed of 25 fps in detecting logos in videos.