<p>Open world object detection is a challenging task that requires models to locate and identify known and potentially unknown objects. Unknown classes are not labeled during the training process, which often leads to the neglect of unknown class objects in traditional detection methods. Due to false positives and missed detections, the current open world object detection methods do not have satisfactory detection performance for unknown objects. Aiming at the issues, we propose a novel unknown-class sensitive open world detector (UCS-OWD) for detecting both known and unknown objects. To prevent unknown objects from being erroneously suppressed, the method constructed an open world objectness score (OWOS), which learns the commonalities of known object characteristics. Besides, OWOS can extend to unknown objects according to the supervised information of known objects. For improving the sensitivity of OWOS to unknown objects, the method further designs a label expansion in the training stage, and the label expansion selects high-quality results of RPN outputs as the supervised information of collaborative real labels. Additionally, a multi scale dilated attention (MSDA) module of the method helps OWOS learn object features better. To better determine the optimal box of unknown objects, the method proposed a distance-sensitive non-maximum suppression (DS-NMS) module with a novel class-adaptive K-means. The proposed class-adaptive K-means can automatically determine the number of the unknown objects by iterative learning, and DS-NMS further constructs the feature similarity matrix and IoU matrix for eliminating redundant prediction boxes. Extensive experiments have shown the superior detection performance in two open world datasets.</p>