Guided Learning for Efficient Optical Flow Estimation

摘要

Current optical flow estimation models require a large number of computing resources, resulting in excessive runtimes to perform inference for mobile devices. Considerable efforts have been exerted to compress deep neural networks, but these approaches have resulted in reduced accuracy. Comparing with other network compression methods, guided learning achieves competitive performance without the need of modifying or retraining the original network as network pruning or specific hardware environment support as network quantization. Guided learning has demonstrated excellent improvements in classification and detection. However, there is still leak of research on guided learning for optical flow estimation. In this work, we propose a novel framework to learn a compact and fast optical flow estimation network, and it can achieve competitive performance through guided learning. We train the small network through additional supervision from high-level feature maps in original optical flow estimation networks to effectively learn robust and efficient features. The overall compression framework is implemented by distilling knowledge from the feature maps of the decoder in original network with a proposed feature guiding loss to transfer knowledge to small network. Lastly, we conduct a comprehensive empirical evaluation over three datasets. Results show consistent improvement in accuracy-speed tradeoffs for the optical flow estimation network.