Open Access

Christian Eggert

• Many companies conduct market research to gather information about their customers. Among other things, companies are often interested in information about customer expectations, customer satisfaction, brand popularity and brand perception. A classical instrument of market research to gather such data are surveys. While surveys undeniably offer valuable insights into many aspects of customer relations, they are time-consuming and expensive to conduct. Because of these limitations, surveys can often be only conducted on a comparatively small sample of people. Therefore, market research organizations are always interested in obtaining data through different channels. The rise of social media has opened up such a channel. With cell phone cameras being ubiquitous nowadays, many people post pictures of their daily activities on social media sites. In doing so, they capture their interactions with certain brands. This capture of brand interaction sometimes is a conscious act, e.g. byMany companies conduct market research to gather information about their customers. Among other things, companies are often interested in information about customer expectations, customer satisfaction, brand popularity and brand perception. A classical instrument of market research to gather such data are surveys. While surveys undeniably offer valuable insights into many aspects of customer relations, they are time-consuming and expensive to conduct. Because of these limitations, surveys can often be only conducted on a comparatively small sample of people. Therefore, market research organizations are always interested in obtaining data through different channels. The rise of social media has opened up such a channel. With cell phone cameras being ubiquitous nowadays, many people post pictures of their daily activities on social media sites. In doing so, they capture their interactions with certain brands. This capture of brand interaction sometimes is a conscious act, e.g. by showing off a new car to friends or by displaying affection to a certain brand of beer. However, more commonly these brand interactions are captured inadvertently. For example, a self portrait might capture the brand of clothing the user likes to wear. In any case, images on social media provide a useful source of information for market research. The reliable detection of company logos forms an important building block for any further analysis of these brand interactions. Object detection is a well-studied problem in computer vision and great advances have been made especially in recent years with the advent of object detection pipelines like R-CNN and Fast(er) R-CNN, which are based on deep convolutional neural networks. In this work, we have applied these new techniques to the problem of company logo detection. We have created a new object detection dataset as a benchmark, specifically targeted for company logo detection. And while on the face of it, company logo detection is nothing but a special case of object detection, we have noticed some peculiarities when applying these new object detection pipelines to our new dataset. We have noticed for example, that while being slower, the oldest approach (R-CNN) performs considerably better on our dataset than newer approaches like Fast(er) R-CNN which are usually considered to be improvements over R-CNN -- both in terms of detection performance and speed. In this work we investigate the reasons for this discrepancy by analyzing each step of these object detection pipelines. In particular, we look at the generation of both heuristic and trainable object proposals and their classification. For heuristic object proposals we look at two commonly used algorithms: Selective Search and Edge Boxes. We observe some conditions under which these algorithms fail that have particular relevance for company logos. In particular we notice that both algorithms struggle to identify proposals for text-based company logos and introduce an additional heuristic to mitigate this problem and demonstrate its effectiveness. For trainable object proposals we look at Region Proposal Networks which we analyze in detail, both theoretically and in in practice and notice some fundamental shortcomings for detecting small company logos. We introduce some simple modifications and show that these are able to considerably improve the performance of small object proposals. We also look at the classification stage and identify the receptive field of the network as an important quantity to improve the classification performance. We finally look at SSD, a more modern single-stage approach for object detection which allows us to incorporate all our observations into a single detection framework. Our improvements allow us to improve our object detection pipeline to a point where where we not only exceed the detection performance of R-CNN but can also perform real-time company logo detection.…