Everything in the Open World: Towards Universal Object Detection”

This work trains the UniDetector on the 115k images from the COCO 2017 training set and evaluates it on the validation set of 5,000 images and reports the closed-world AP on the test-dev split with 20k images.