DeepLab is a state-of-art deep learning model for semantic image segmentation. For tensorflow implemention, please refer to [tensorflow/models/research]: https://github.com/tensorflow/models/tree/master/research/deeplab
- Encoder-Decoder Structures
- Atrous Convolutions in ASPP structure
- Separable Convolutions instead of traditional Convolutions
- Bilinear Interpolation instead of Transpose Convolutions
Encoder
- Xception as the Feature Extractor
- ASPP
- Separable Convolutions
Decoder
- Upsampling
- Separable Convolutions