Modern object detection models often require enormous training images with accurate annotations for each scenario; it is a significant obstacle for actual applications for computer vision. In this paper, we propose a novel framework for training lightweight object detection models without additional manual annotations by inheriting the rich expression power of multiple pre-trained visual-language(VL) models. The key is to obtain elaborate pseudo labels for lightweight model by knowledge-extraction training from multiple VL models, the biases of which are corrected by score correction. We can obtain accurate detection labels without using any prior manual annotations for each image by using novel data augmentation to enhance knowledge extraction from the VL models and pseudo-label integration. In contrast to current semi-supervised and unsupervised approaches for object detection, our proposed framework is immediately applicable to state-of-the-art object detection models and training protocols. Comprehensive experiments on two public datasets demonstrated that our framework is fast and lightweight while maintaining accuracy, surpass supervised models.