Learning Open-World Object Proposals Without Learning to Classify

Object proposals have become an integral pre-processing step of many vision pipelines including object detection, weakly supervised discovery, tracking, etc. Compared to the learning-free methods, learning-based popular recently due growing interest in detection. The common paradigm is learn from data labeled with a set regions and their corresponding categories. However, this approach often struggles novel objects open world that are absent training set. In letter, we identify problem binary classifiers existing proposal methods tend overfit Therefore, propose classification-free Localization Network (OLN) which estimates objectness each region purely by how well location shape overlap any ground-truth (e.g., centerness IoU). This strategy learns generalizable outperforms on cross-category generalization COCO. We further explore more challenging cross-dataset onto RoboNet EpicKitchens dataset, long-tail detection LVIS dataset. demonstrate clear improvement over state-of-the-art detectors proposers. code publicly available at https://github.com/mcahny/object_localization_network .