Revisiting Open World Object Detection [arXiv]
See INSTALL.md.
Our new data division is based on COCO2017. We divide the training set into four tasks, in which each task has 20 categories. For each task, we obtained images containing the categories of each task from the training set, and removed the annotation information of other categories in these pictures during training. In each task, 1000 images are sampled as the validation set. And we de duplicate the training set and the validation set. For the testing set, we adopt the validation set of COCO2017, which contains relatively complete annotation information.
The data files are at ./datasets/Main/.