Human action recognition in videos is a challenging task in the field of computer vision. Based on the idea of integrating temporal and spatial feature, many works have proposed a variety of methods for extracting spatiotemporal features, such as two-stream network and 3D convolution neural network (3D-CNN). However, due to the huge computational cost of optical flow for two-stream network and the huge number of parameters of 3D-CNN, the computational time required for action recognition is very long, therefore it is difficult to meet the requirements of real-time recognition. This paper aims to explore an efficient architecture of 3D-CNN for action recognition. On the premise of guaranteeing the recognition accuracy, we aim to greatly reduce the computational cost. In order to ensure good performance while reducing the amount of input data, we present Global Evaluate-and-Rescale (GER) Network, which is able to automatically extract the key frames of input data. We have evaluated the performance of our proposed model on two challenging human action recognition datasets UCF101 and HMDB51. The experimental results show that GER Network can reduce up to 50% of the computation time for recognition while achieving approximate accuracy with state-of-the-art 3D-CNN models.