Dynamic gesture recognition, which plays an essential role in human-computer interaction, has been widely investigated but not yet fully addressed. The challenge mainly lies in three folders: 1) to model both of the spatial appearance and the temporal evolution simultaneously; 2) to address the interference from the varied and complex background; 3) the requirement of real-time processing. In this paper, we address the above challenges by proposing a novel deep deformable 3D convolutional neural network for end-to-end learning, which not only gains impressive accuracy in challenging datasets but also can meet the requirement of the real-time processing. We propose three types of very deep 3D CNNs for gesture recognition, which can directly model the spatiotemporal information with their inherent hierarchical structure. To eliminate the background interference, a light-weight spatiotemporal deformable convolutional module is specially designed to augment the spatiotemporal sampling locations of the 3D convolution by learning additional offsets according to the preceding feature map. It can not only diversify the shape of the convolution kernel to better fit the appearance of the hands and arms, but also help the models pay more attention to the discriminative frames in the video sequence. The proposed method is evaluated on three challenging datasets, EgoGesture, Jester and Chalearn-IsoGD, and achieves the state-of-the-art performance on all of them. Our model ranked first on Jester's official leader-board until the submission time. The code and the trained models are released for better communication and future works(1). (C) 2020 Elsevier Ltd. All rights reserved.