Deep Learning - Deeplab Series Theories

Motivation

UNet has restricted receptive fields. It is sufficient for identifying local areas such as tumors (in medical images). When we learn larger image patches, UNet was not sufficient.

DeepLab V1 & V2 (2016) [1]: they were reviewed together as they both use Atrous convolution, or “dilated convolution” (空洞卷积), and Fully Connected Conditional Random Field. V1 was only using VGG16, while V2 was using VGG16 and ResNet. DeepLab V3 (2017) [2] DeepLab v3+ (2018) - Achieved SOTA on Pascal-VOC 2012.

The Use of Dilated Convolution

• Some notes about Dilated Convolution is here

• An interesting question is: why don’t we use a larger kernel? Since VGG networks (2014), using a 3x3 kernel has become a trend. That’s because:
  • A 7x7 is more than 5 times larger than a 3x3 kernel.
  • A 3x3 introduces more non-linearity than a 7x7, which allows the network to learn more complex landscapes.
• Fun fact: after 2014, 3x3 seems to become the most popular kernel size

ASPP

TODO

References

[1] He, K., Zhang, X., Ren, S., & Sun, J. (2015). Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 37(9), 1904–1916.

[2] Chen, L. C., Papandreou, G., Kokkinos, I., Murphy, K., & Yuille, A. L. (2016). DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(4), 834–848. https://doi.org/10.1109/TPAMI.2017.2699184

[3] Chen, L. C., Papandreou, G., Schroff, F., & Adam, H. (2017). Rethinking Atrous Convolution for Semantic Image Segmentation. arXiv preprint arXiv:1706.05587.

[3] Chen, L. C., Zhu, Y., Papandreou, G., Schroff, F., & Adam, H. (2018). Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation. Proceedings of the European Conference on Computer Vision (ECCV), 801–818.

• Previous
  Deep Learning - Spatial Pyramid Pooling
• Next
  Deep Learning - MobilenetV2 Hands-On