Page - 92 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

Confusing Similarity between Visual Trademarks A Dataset Based on USTTAB Examinations* Lukas Knoch1 and Mathias Lux2 Abstract—Trademarks are an important visual clue for customers to identify brands, products and companies, and can influence the buying decision significantly. One major problem with visual trademarks is, that newly registered trademarks are required by law not to be visually similar to existing ones. Therefore, automatic detection of visually similar trademarks is an important use case for content based image retrieval. Confusing similarity between trademarks is defined by law, and numerous cases of theUnitedStatesTrademarkTrial andAppeal Board (USTTAB) handling trademark similarity are available. In this paper we present a novel and freely available data set for evaluation of trademark similarity algorithms based on real life data, ie. all registered trademarks in the USA as well as USTTAB decisions and expert opinions. The data set should serve as a basis for further investigations, ie. extension of the data set by crowd sourcing and consideration of the intuitive concept of visually confusing similarity. I. INTRODUCTION Visual trademarks, or logos, often influence our buying decisions and are therefore valuable goods for the companies owning the visual trademark. A common and well known example is the Apple company logo (compare Figure I) present on iPhones, iPads and Apple computers. Apple Computers invests time and money to find out if other companies worldwide use similar logos on similar products. The same approach is also taken by many companies who define themselves through their brands, like Nike, Adidas, or Red Bull. Fig. 1. Examples of well known logos and protected trademarks in many countries including the Apple logo, the Github logo and the Nike swoosh. To avoid confusion between different trademarks, they must be dissimilar enough to each other. Some companies even try to trick customers by deliberately using trademarks * This article is based on the master’s thesis of Lukas Knoch and has been done with the support of the World Intellectual Property Organization, WIPO, partially as the result of an internship at the UN Headquarters in Geneva, CH 1 Lukas Knoch was student at Alpen-Adria Universita¨t Klagenfurt, Austrialukas.knoch@aau.at 2 Mathias Lux is Associate Professor at the Institute for Information Technology at Alpen-Adria Universita¨t Klagenfurt mathias.lux@aau.at that are similar to well known signs. To avoid fraud, trade- marks can be protected by law. There are several offices in charge of managing trademark registrations for different regions including the European Union Intellectual Property Office (formerly Office for Harmonization in the Internal Market, short OHIM) or the United States Patent and Trade- mark Office (short USPTO). If a new trademark is registered, it has to be ensured that there is no confusing similarity to any other previously registered marks. This difficult job is executed by professional trademark examiners who compare the different trademarks to each other and decide about the similarity. While there are systems in place like the textual Vienna Classification [21], taxonomies which are intended to help the examiner, these systems are tedious and error prone as they rely on manual annotation. Another way of assisting the examiners are visual trade- mark retrieval systems. These systems can take a specific trademark as an input and deliver a set of trademarks ranked bysimilarity to thequery image,which iscommonly referred to as query by example in content based image retrieval. While several systems have been proposed [28], [9], [15], their retrieval performance leaves a lot of room for improve- ment [25]. There are several papers suggesting new algo- rithms for visual trademark retrieval, but their evaluations are based on trademark datasets downloaded from the internet [22], [23], pure shape datasets like MPEG-7 [13], [1] or hand picked ground truth [27], [20], [5], [26]. Unfortunately, objective evaluation of these systems is currently hardly possible as there are no datasets available that (i) represent real world data, ie. the actual visual trademarks registered at the trademark offices, and (ii) that are based on expert opinions and court decisions. To aid with the development of content based visual trade- mark retrieval systems, this paper introduces a realistic novel dataset based on real world trademark trials. Our dataset can provide the base for research on content based visual information retrieval systems.Thedatasetcontains1,859,218 visual trademarks registered at the United States Patent Office (USPTO) as well as three different sets of ground truths based on trials at the United States Trademark Trial and Appeal Board (USTTAB). The raw visual trademarks and trial data is provided by Google12, the extracted meta data is available at a public website3. 1https://www.google.com/googlebooks/ uspto-trademarks-usamark.html, last visited 2016-01-19 2https://www.google.com/googlebooks/ uspto-trademarks-ttab.html, last visited 2016-01-19 3www.rumpelcoders.at/usttabdataset 92