Traditional object detectors usually require fully annotated instances for training to achieve satisfactory detection results. However, for incompletely annotated image instances, especially in medical imaging, their detection results are still far from satisfactory due to the difficulty for sample annotation. In this regard, Co-mining self-supervised learning for incompletely annotated object detection performs excellently. However, its effectiveness in detecting smaller objects is not yet satisfactory. Therefore, this paper proposes an improved model called CO-FCOS to design and implement a better incompletely annotated detection algorithm for small objects. In order to fully utilize the position and category information of existing objects in the image, more effective data augmentation strategies are applied, and the predictions are generated by detectors in two branches respectively. To further improve the model’s adaptability and generalization performance under different scenarios and data distributions, an OTA dynamic label assignment strategy is introduced. In addition, the label consistency strategy breaks the limitations of the original model, generates more pseudo labels for the model, which enriches the training data, and improves the model performance. Experimental results show that, compared to the Co-mining model, our proposed CO-FCOS model improves the accuracy by 4.3 mAP on the cervical cancer TCT dataset and by 1.7 mAP on the COCO-miss50 incomplete annotation dataset.