The Flickr Logos 27 dataset is an annotated logo dataset downloaded from Flickr and contains more than four thousand classes in total. It consists of three image collections/sets.
The training set contains 810 annotated images, corresponding to 27 logo classes/brands (30 images for each class). All images are annotated with bounding boxes of the logo instances in the image. We allow multiple logo instances per class image. The training set is randomly split in six subsets, each one containing five images per class.
The distractor set contains 4207 logo images/classes, that depict, in most cases, clean logos. All images come from the Flickr group Identity + Logo Design. Each one of the distractor set images defines its own logo class and we regard the whole image as bounding box.
Finally, the query set consists of 270 images. There are five images for each of the 27 annotated classes, summing up to 135 images that contain logos. Furthermore, the query set contains 135 Flickr images that do not depict any logo class, giving 270 test images in total.
The brands included in the dataset are: Adidas, Apple, BMW, Citroen, Coca Cola, DHL, Fedex, Ferrari, Ford, Google, Heineken, HP, McDonalds, Mini, Nbc, Nike, Pepsi, Porsche, Puma, Red Bull, Sprite, Starbucks, Intel, Texaco, Unisef, Vodafone and Yahoo.
If you use the Flickr Logos dataset please cite our ICMR paper shown below.
For each image of the annotated set, the tab separated text file below contains one line with the following fields:
For the 4K distractor logo images, as bounding box we regard the whole image and each one of them defines its own class. Thus, we do not provide a bounding box.
For each image of the query set, the tab separated text file below contains one line with the following fields:
For the training and query sets we provide both the image files and the annotation text files in the format explained above. For the distractor set we just provide the Flickr medium URLs and the images can be downloaded directly from Flickr.
There is a similar dataset called FlickrLogos-32. It consists of 32 classes and while there is some overlap it is largely different from our Flickr Logos 27 dataset. Results for these different datasets are not comparable.