Context

We created this dataset to provide us a balanced multiclass dataset for training and testing our competition models before submissions to Google Lanmark Recognition 2020 competition.

Content

The dataset contains 98k images and their corresponding image ids and labels. Those image ids are specifically chosen since they have exactly 200 samples of them. The image features are extracted by the globally averaged topless VGG-16 network with Imagenet weights (outputs 512 features) and the globally averaged topless ResNet-50 network with Imagenet weights (outputs 2048 features).

Acknowledgements

I want to thank to the collaborators of the dataset.