Tensorflow-PSPNet

This is an improvement of the implementation of pudae's. The improvements include:

1. Support ResNet-101, which is the original backbone network of hszhao's PSPNet.
2. Support auxiliary loss.
3. Support weighted softmax loss.
4. Convert the pspnet101_VOC2012.caffemodel to Tensorflow model, which can be downloaded from Baidu Yun Pan or Goole Drive. Download it and you can test VOC2012's images with test_segmentation.sh.

The way to convert the caffe model to TF model can be done with caffe-tensorflow. With an ugly way I converted the caffe model's names to the name scopes of PSPNet defined in pspnet_v1.py. It only supports pspnet101_VOC2012.caffemodel. For other networks there might be some slight modifications, such as the size of the input image.

Prepare your dataset

Create your dataset description according to ade20k.py and register it in dataset_factory.py. Run convert_data.sh to convert your dataset to tfrecord.

Training

To train your dataset, modify the parameters in train_pspnet.sh. Set DATASET_NAME according to your needs and run the scrpit:

./train_pspnet.sh

Evaluation

Modify eval_pspnet.sh and run it:

./eval_pspnet.sh

Inference

Set the image your want to test in test_segmentation.sh and run:

./test_segmentation.sh