Video Semantic Segmentation With Distortion-Aware Feature Correction

Video semantic segmentation is active in recent years benefited from the great progress of image segmentation. For such a task, per-frame generally unacceptable practice due to high computation cost. To tackle this issue, many works use flow-based feature propagation reuse features previous frames. However, optical flow estimation inevitably suffers inaccuracy and then causes propagated distorted. In paper, we propose distortion-aware correction alleviate which improves video performance by correcting distorted features. be specific, firstly transfer distortion patterns into space conduct effective map prediction. Benefited guidance maps, proposed Feature Correction Module (FCM) rectify areas. Our method can significantly boost accuracy at low price. The extensive experimental results on Cityscapes CamVid show that our outperforms state-of-the-art methods.