Cut and Learn for Unsupervised Object Detection and Instance Segmentation

OpenImages Datasets AP 50 Figure 1 . Zero-shot unsupervised object detection and instance segmentation using our CutLER model, which is trained without human supervision. We evaluate the model using the standard detection AP box 50 . CutLER gives a strong performance on a variety of benchmarks spanning diverse image domains -video frames, paintings, clip arts, complex scenes, etc. Compared to the previous stateof-the-art method, FreeSOLO [47] with a backbone of ResNet101, CutLER with a backbone of ResNet50 provides strong gains on all benchmarks, increasing performance by more than 2× on 10 of the 11 benchmarks. We evaluate [47] with its official code and checkpoint.