DenseASPP for Semantic Segmentation in Street Scenes

1. DenseASPP for Semantic Segmentation in Street Scenes CVPR 2018 Maoke Yang, Kun Yu, Chi Zhang*, Zhiwei Li*, Kuiyuan Yang* DeepMotion .Inc *MSRA Speaker: Zhixuan Li

2. Important challenge in Scene Segmentation • Scale variation • People • Cars

3. Importance of Large Reception Field Big Object Small Kernel Big Object Big Kernel Small Object Small Kernel

4. How to enlarge Reception Field • Atrous convolution  • Atrous convolution is proposed to enlarge the size of reception field and dicrease the cost of computing simultaneously.

5. How to enlarge Reception Field • ASPP • ASPP is proposed to concatenate feature maps generated by atrous convolution with different dilation rates.

6. Disadvantage of ASPP When training high resolution images, the size of reception field will never become big enough. And large enough dilation ratio (d>24) will decrease the power of model.

7. DenseASPP Layer cascade Through cascade, applying d=6 conv on the feature map of d=3 conv, we can actually get the bigger receptive field than a normal d=6 conv can get.

8. Network Structure

9. Rank of the experiment results

10. Conclusion Cascadeis an important thought, separating a big problem to some simple problems is very helpful sometimes.