This repository holds a keras porting of the deeplabV2 model presented in this paper.

Please note that only the VGG encoder version with voc pretraining on VOC12 segmentation is held in this repo.

The model is developed and tested with Theano backend. In order to get it working with Tensorflow, you'll have to code the BilinearUpsampling layer, and convert all weights.

Moreover, this network does not include the CRF post processing, but it can be integrated with this code.

Simply get the model with the function DeeplabV2 in deeplabv2.py. For an example, check predict.py.

• add model with ResNet-101 encoder.