by Hang Su,Shaogang Gong,Xiatian ZhuUnknown
by Hang Su,Shaogang Gong,Xiatian ZhuLicense : Unknown
The WebLogo-2M dataset is a weakly labelled (at image level rather than object bounding box level) logo detection dataset. The dataset was constructed automatically by sampling the Twitter stream data. It contains 194 unique logo classes and over 2 million logo images. It features with large scale but very noisy labels across logos due to the inherent nature of web data. Generally, these weakly labelled logo images are used for model training. Therefore, this dataset is designed for large-scale logo detection model learning from noisy training data with high computational challenges. For performance evaluation, we further provide 6, 569 test images with manually labelled logo bounding boxes for all the 194 logo classes.