In stereo vision, depth information relies on the dense registration accuracy of binocular stereo images, and its realtime performance is also significant in many automation applications. Recently, it is still a challenge to balance the efficiency and accuracy. Motivated by this problem, we propose a lightweight 2D guided deformable aggregation(GDA) module. It uses color prior information to learn the aggregation sampling points for fitting the irregular window. And it enables to fast recover the lost high-frequency detail information from a coarse cost volume. Furthermore, we propose a guided deformable aggregation based stereo matching network (GDANet) for balancing the efficiency and accuracy. It builds a fast 3D network to obtain the cost volume of low-frequency non-detail regions, and then uses the lightweight 2D GDA module to recover high-frequency detail regions. Experiments show that GDANet achieves better results than current high efficiency methods in SceneFlow and KITTI datasets. Especially, in edge regions and thin structures, our method shows better qualitative and quantitative results.