NPU-TM: A web image dataset for similar trademark search

The NPU-TM (V1.0) was constructed by collecting web images and labelled by the humans. It contains 7139 trademark images, in which 642 groups of visually similar images are identified and the rest are totally different from previous ones. Each group has about 13.8 images averagely in the NPU-TM dataset and 52 images maximum in a group. NPU-TM has sufficient training samples for learning parameters of deep model as it contains many similar groups.