Spatio-Temporal Recurrent Networks for Event-Based Optical Flow Estimation

Event camera has offered promising alternative for visual perception, especially in high speed and dynamic range scenes. Recently, many deep learning methods have shown great success providing model-free solutions to event-based problems, such as optical flow estimation. However, existing did not address the importance of temporal information well from perspective architecture design cannot effectively extract spatio-temporal features. Another line research that utilizes Spiking Neural Network suffers training issues deeper architecture. To these points, a novel input representation is proposed captures events distribution signal enhancement. Moreover, we introduce recurrent encoding-decoding neural network estimation, which Convolutional Gated Recurrent Units feature maps series event images. Besides, our allows some traditional frame-based core modules, correlation layer iterative residual refine scheme, be incorporated. The end-to-end trained with self-supervised on Multi-Vehicle Stereo Camera dataset. We it outperforms all state-of-the-art by large margin.