Recently, semi-supervised learning attracts extensive interests in the field of object detection, since it is beneficial for alleviating the label annotation burden. Existing methods generates unsatisfactory pseudo labels and have imperfect performance on the detection of small objects. In this paper, we propose a novel semi-supervised learning algorithm for object detection, based on cross-scale siamese representation learning. This helps improve the generalization of learned features, thus reducing the dependence to high-quality annotations. Furthermore, we devise box refinement and adaptive thresholding strategies to generate higher-quality pseudo labels for unlabeled images. Experiments on VOC and COCO demonstrate that our method achieves new state of the art. Specially, it brings gain of 1 mAP on COCO with 10% labeled images.