CCNet CrissCross Attention for Semantic Segmentation Zilong Huang

CCNet: Criss-Cross Attention for Semantic Segmentation Zilong Huang 1, Xinggang Wang 1, Lichao Huang 2, Chang Huang 2, Yunchao Wei 3, Wenyu Liu 1 1 School of EIC, Huazhong University of Science and Technology 2 Horizon Robotics 3 Beckman Institute, University of Illinois at Urbana-Champaign Speaker: Zhixuan Li 2019/03/11 Zhixuan Li 1

Retrospect DANet 2019/03/11 Zhixuan Li 2

Retrospect DANet • Computation Cost Expensive: • Each pixel has to multiply with all pixels in the feature map 2019/03/11 Zhixuan Li 3

About CCNet • CCNet is the computation cost saving version of DANet • RCCA Module 2019/03/11 Zhixuan Li 4

Two RCCA modules are necessary • A simple example, after the first RCCA module, the value in position C contains the information of A and D. • And in the same time, the value in A contains the information of B and C. • So after the second RCCA module, position C can capture the relationship between C and B indirectly by computing the distance with the new value in position A. 2019/03/11 Zhixuan Li 5

Ablation studies • The effect of attention module 2019/03/11 Zhixuan Li 6

Ablation studies • Comparison of context aggregation approaches on Cityscapes validation set • • PSP : Pyramid Spatial Pooling ASPP : Atrous PSP NL : DANet RCCA : CCNet 2019/03/11 Zhixuan Li 7

Rank of the experiment result 2019/03/11 Zhixuan Li 8

Visualization of Attention Map 2019/03/11 Zhixuan Li 9

Conclusion • Advantages • The decomposition greatly reduce the complexity in time and space from O((H x W)) to O((H x W)(H + W - 1)). 2019/03/11 Zhixuan Li 10

Conclusion • Disadvantages • At the second RCCA module, when computing C, A is used twice but B is used just once when computing A. • So maybe A makes more affect to C than B does. • It's unreasonable for B maybe more important for C than A. 2019/03/11 Zhixuan Li 11