Temporally Consistent Depth Prediction with Flow-Guided Memory Units.

摘要: Predicting depth from a monocular video sequence is an important task for autonomous driving. Although it has advanced considerably in the past few years, recent methods based on convolutional neural networks (CNNs) discard temporal coherence and estimate independently each frame, which often leads to undesired inconsistent results over time. To address this problem, we propose memorize consistency sequence, leverage of prediction. end, introduce two-stream CNN with flow-guided memory module, where stream encodes visual features, respectively. The implemented using gated recurrent units (ConvGRUs), inputs features sequentially together optical flow tailored our task. It memorizes trajectories individual selectively propagates spatial information time, enforcing long-term prediction results. We evaluate method KITTI benchmark dataset terms accuracy, runtime, achieve new state art. also provide extensive experimental analysis, clearly demonstrating effectiveness approach memorizing