Semantic Video CNNs through Representation Warping

Semantic Video CNNs through Representation Warping

來自專欄 計算機視覺與深度學習

17年的ICCV.

這篇文章提出了對視頻分割通用的網路module, 叫做NetWarp, 使用快速的optical flow設計了module結構, 來提高準確率並且可以end-to-end train.

NetWarp的結構:

主要的步驟有:

1.flow Computation(使用DIS-Flow)

2.Flow Transformation(We concatenate the original two channel flow, the previous and present frame images, and the difference of the two frames. This results in a 11 channel tensor as an input to the FlowCNN.)

3.Warping Representations

4.Combination of Representations

最後的結構在cityscapes上也是提高了1個多點:

推薦閱讀:

TAG:計算機視覺 | 深度學習DeepLearning |