QMUL OpenLogo

Description

Existing logo detection benchmarks consider artificial deployment scenarios by assuming that large training data with fine-grained bounding box annotations for each class are available for model training. Such assumptions are often invalid in realistic logo detection scenarios where new logo classes come progressively and require to be detected with little or none budget for exhaustively labelling fine-grained training data for every new class. Existing benchmarks are thus unable to evaluate the true performance of a logo detection method in realistic and open deployments. In this work, we introduce a more realistic and challenging logo detection setting, called Open Logo Detection. Specifically, this new setting assumes fine-grained labelling only on a small proportion of logo classes whilst the remaining classes have no labelled training data to simulate the open deployment. Further, we create an open logo detection benchmark, called QMUL-OpenLogo, to promote the investigation of this new challenge. QMUL-OpenLogo contains 27,083 images from 352 logo classes, built by aggregating and refining 7 existing datasets and establishing an open logo detection evaluation protocol.

Statistics

For dataset training-evaluation split, we propose 3 kind of setting. The first one is the fully supervised setting, with every logo classes contains 70% of training split and 30% of evaluation split. The second setting split the dataset into 32 supervised classes and 320 unsupervised classes, where the supervised contains real training split and evaluation split, while the unsupervised classes have no training split, only evaluation split. For the 3rd setting, we split the dataset into 176 supervised classes and 176 unsupervised classes.

Notification for Supervised/Unsupervised splitting: The dataset contains images of multi-instances and multi-classes, thus some supervised set images may contain both supervised and unsupervised instances in it, it is recommendded to filter it for own useage.

Closed-set logo detection setting: all logo classes have training images.

Classes Train Images Val Images Test Images Total Images
352 15,975(59.0%) 2,777(10.3%) 8,331(30.7%) 27,083

Open-set logo detection setting 1: only a msall proportion of logo classes have training images.

Type Classes Train Images Val Images Test Images Total Images
Supervised 32 1,280(11.2%) 1,023(8.9%) 9,168(79.9%) 11,471
Unsupervised 320 0 0 15,612 15,612
Total 352 1,280(4.7%) 1,023(3.8%) 24,780(91.5%) 27,083

Open-set logo detection setting 2: Half of the logo classes have training images and the other half have no training images.

Type Classes Train Images Val Images Test Images Total Images
Supervised 176 10,586(59.3%) 1,561(10.2%) 3,121(20.5%) 15,268
Unsupervised 176 0 0 3,649 3,649
Total 352 10,586(56.0%) 1,561(8.2%) 6,770(35.8%) 18,917

News

  • March 4, 2019: Baidu cloud download link provided
  • Feb 19, 2019: Tencent cloud download link provided
  • Nov 7, 2018: Leaderboard introduced
  • July 03, 2018: Dataset released.

Download

QMUL-OpenLogo Dataset (4.7 GB): [Google Drive] [Baidu Cloud] [Tencent cloud]

Supplyment Imageset List for OpenLogo Semi-Supervised Challenge Setting 1: [Google Drive]

Supplyment Imageset List for OpenLogo Semi-Supervised Challenge Setting 2: [Google Drive]

Clean Logo Exemplars: [Google Drive]

Citation

	    	Open Logo Detection Challenge.
		Hang Su, Xiatian Zhu and Shaogang Gong.
		In Proc. British Machine Vision Conference (BMVC), Newcastle, UK, September 2018.
		BibTex arxiv
		
	    

Licence

Please notice that, the QMUL-OpenLogo Dataset is made available for academic research purpose only. All the images were collected from the existing logo detection datasets, and the copyright belongs to the original owners.

Contact

Please feel free to send questions/comments/results(to be added in Leaderboard) to Hang Su at hang.su@qmul.ac.uk