Traditional Algorithms Never Die As They Can Find Their Position In Deep Neural Networks

[1] Liu, S., De Mello, S., Gu, J., Zhong, G., Yang, M. H., & Kautz, J. (2017). Learning affinity via spatial propagation networks. In NIPS2017.

[2] Pan, X., Shi, J., Luo, P., Wang, X., & Tang, X. (2017). Spatial As Deep: Spatial CNN for Traffic Scene Understanding. In AAAI2018.

[3] Conditional Random Fields as Recurrent Neural Networks

[4] Semantic Image Segmentation with Deep Convolutional Nets and Fully Connected CRFs

[5] Efficient Inference in Fully Connected CRFs with Gaussian Edge Potentials

 

Neural Network, CNN in particular, is not omnipotent in many computer vision tasks by itself, thus the need for traditional algorithms such as CRF arises.

Take Instance or semantic segmantation for example,  popular methods such as FCN are giving coarse results, but if we apply an extra step using CRF to 

refine the result given by FCN, we can improve the result to a whole new level.

 

TBA