Fault interpretation is an important part of seismic structural interpretation and reservoir characterization. In the conventional approach, faults are detected as reflection discontinuity or abruption and are manually tracked in post-stack seismic data, which is time-consuming. In order to improve efficiency, a variety of automatic fault detection methods have been proposed, among which widespread attention has been given to deep learning-based methods. However, deep learning techniques require a large amount of marked seismic samples as a training dataset. Although the amount of synthetic seismic data can be guaranteed and the labels are accurate, the difference between synthetic data and real data still exists. To overcome this drawback, we apply a transfer learning strategy to improve the performance of automatic fault detection by deep learning methods. We first pre-train a deep neural network with synthetic seismic data. Then we retrain the network with real seismic samples. We use a random sample consensus (RANSAC) method to obtain real seismic samples and generate corresponding labels automatically. Three real 3D examples are included to demonstrate that the fault detection accuracy of the pre-trained network models can be greatly improved by retraining the network with a few amount of real seismic samples.