Knowledge transfer for Object Detection with Evolution architecture search

Deep learning has been proved to achieve excellent results in various fields, and appropriate network architecture and sufficient data play an important role. Due to the high cost of annotation for the task of object detection, domain adaptation methods have been introduced in this field. But these methods are based on the rigid network architecture and bound to the input dimension of the adaptive module, which is not only difficult to better balance accuracy and speed of detection model, but also can not use the multi-scale training method, resulting in the reduction of the application scenario of the model. Inspired by this problem, we propose a new object detection method based on multi-scale adversarial domain adaptation and network architecture search. An evolutionary algorithm is adopted to help search the network architecture to balance accuracy and speed. The ability of domain adaptation can also be effectively improved by the searched architecture. The experimental results have demonstrated the significant improvement that benefited from the framework in terms of its performance and computational efficiency in solving unlabeled object detection.