Logos in the Wild dataset

<p>Logos in the Wild dataset - unofficial version</p> <p>I am uploading the following dataset even though I am not the original author (see https://arxiv.org/abs/1710.10891). I built an unofficial implementation of the algorithm using this dataset (named Logohunter) and have received emails of interested researchers looking for the dataset; unfortunately the original IOSB Fraunhofer link referred in the paper is dead. In addition, as the original dataset only provides URLs to the images (some of which have disappeared with time), I am uploading the images themselves as well. While copyrighted, this is clearly fair use.</p> <p>The <span class="ng-binding">LogosInTheWild-v2.zip file contains the original dataset (URLs only), while </span><span class="ng-binding">litw_cleaned.tar.gz&nbsp;</span> contains the full dataset with images (as downloaded in February 2019).</p> <p>Below is the content of the README from the original dataset authors:</p> <p><br> # General remarks<br> This dataset consists of web images which were crawled via Google<br> image search and according logo annotions. It was collected at<br> Fraunhofer IOSB in Karlsruhe, Germany.<br> For dataset related matters please contact Christian Herrmann:<br> christian.herrmann@iosb.fraunhofer.de.<br> <br> <br> # Structure<br> Each folder contains the raw Pascal VOC style xml annotation files and a<br> urls.txt file containing a list of URLs where the images can be<br> downloaded. Each row in the list contains the image ID and the URL<br> of the image file.<br> <br> A folder includes all images resulting from the Google image search for<br> this brand. Because images can show a large variety of logos beyond the<br> keyword search, there are a lot logos of different brands within each<br> folder or sometimes even within a single image.<br> The bounding box name denotes the actual brand for each logo. When<br> necessary, separation is made between graphical and textual logos via<br> additional specifiers of the brand name (e.g. &#39;porsche-logo&#39;,<br> &#39;porsche-text&#39;). Visually different logos of one brand are separated by<br> enumeration if distinction by graphical/textual is impossible (e.g.<br> &#39;adidas1&#39;, &#39;adidas2&#39;). There are some misspellings and inconsistencies<br> with the labels and specifiers in the raw annotation files. We opt not to<br> alter the raw files provided by the annotation crew but instead fix the<br> issues by the create_clean_dataset.py script (see Scripts section below).<br> <br> List of cleaned specifiers:<br> - &#39;text&#39;: pure textual logo<br> - &#39;symbol&#39;: graphical logo<br> <br> Additional specifiers in raw annotations:<br> - &#39;partial&#39;,&#39;teilsichtbar&#39;: logo is significantly occluded and thus only<br> partially visible, this information is only included in the raw annotations<br> - &#39;schrift&#39;,&#39;schriftzug&#39;: same as &#39;text&#39;<br> - &#39;logo&#39;: same as &#39;symbol&#39;<br> <br> <br> # Scripts<br> To ease processing, the scripts folder contains a Python scripts to<br> preprocess the dataset.<br> create_clean_dataset.py corrects labeling mistakes and can create<br> different versions of the dataset:<br> <br> 1.) Clean Pascal VOC dataset structure which is straight-forward readable<br> by a lot of object detector frameworks. This is created in all cases:<br> python create_clean_dataset.py --in ./data --out ./cleaned-data<br> <br> 2.) Cropped logos sorted into seperate brand folders. This addresses<br> classification or verification tasks. Parameter: --roi.<br> <br> 3.) Logo classes from FlickrLogos32 can be excluded from 1) and 2) via<br> --wofl32. This allows training on Logos in the Wild and testing on<br> FlickrLogos32 if brand overlap is undesired, such as for open-set<br> evalutation.<br> <br> <br> # How to get started<br> 1.) Download the images from the provided URLs.<br> 2.) Execute create_clean_dataset.py script.<br> <br> # Dataset usage<br> If you use this dataset in your work please cite:</p> <p>```<br> @INPROCEEDINGS{,<br> author = {T{\&quot;u}zk{\&quot;o}, Andras and Herrmann, Christian and Manger, Daniel<br> and J{\&quot;u}rgen Beyerer},<br> title = {{O}pen {S}et {L}ogo {D}etection and {R}etrieval},<br> booktitle = {Proceedings of the 13th International Joint Conference on<br> Computer Vision, Imaging and Computer Graphics Theory and Applications:<br> VISAPP},<br> year = {2018}}<br> ```</p>