Detecting instances of object categories unseen by the model during training has essential applications in several downstream computer vision tasks. However, conventional object detectors are not designed to detect unseen, open-world object categories. The performance on unseen categories may be improved by improving the model’s generic objectness criteria, which tends to overfit the seen categories. We introduce a self-learned latent keypoints (Self-LKP)-based training approach that guides a conventional object detection model to learn class-agnostic weights. Specifically, our method allows the model to incorporate objectness cues by learning latent keypoints within each object boundary. This helps improve the model’s ability to detect and localize a generic object. We extensively evaluate the proposed approach for cross-category and cross-dataset setups using three well-established object detection datasets using both one and two-stage detectors. The proposed Self-LKP approach improves average recall by 7.2% on unseen object categories using a one-stage object detector.