Last updated 03/28/2018
Lian qing (lianqinglalala@gmail.com)
- Learning Transferable Architectures for Scalable Image Recognition [pdf]
[code]
- Barret Zoph, Vijay Vasudevan, Jonathon Shlens, Quoc V. Le
Use NASNet's architecture and do some improvement
- use block structure(like Inception or resnet) and only search the block to improve searching time
- improve searching process that first search the network on a small dataset (cifar) and then use that network to test on larger dataset(Imagenet)
- make a comparison between random search and use reinforcement learning and show the benefit from reinforcement learning
- Do a experiment that transfer the network from classification task to detection task
-
rethinking Atrous Convolution for Semantic Image Segmentation [pdf]
- Liang-Chieh Chen, George Papandreou, Florian Schroff, Hartwig Adam
Add Atrous Spatial Pyramid Pooling on the end of the model a) different rate's atrous and convolution b)image level feature Multi-grid Method Once reduce stride one time ,every subsequent conv's atrous rate should time 2 Test if the network can go deeper with resent50 and resnet101 which show slightly improve.
-
Pyramid Scene Parsing Network [pdf][code]
- Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, Jiaya Jia
On the last layer add a Pyramid network with multi brach which first downsampling, then conv to extract feature or context information and then upsamping to the same size, finally concat them