This code implements the model described in Referring Image Segmentation via Recurrent Refinement Networks, CVPR 2018.
This code derives from TF-phrasecut-public.
Please follow its Setup
and Data Preparation
sections except that the DeepLab-ResNet backbone
comes from tensorflow-deeplab-resnet. The
pre-trained DeepLab-ResNet model can be downloaded from
here.
Before training, make sure refer
, cocoapi
, and deeplab
are in PYTHONPATH
.
export PYTHONPATH=./external/refer:./external/cocoapi/PythonAPI:./external/tensorflow-deeplab-resnet:$PYTHONPATH
To train a model on UNC dataset, run
python main_convlstm_p543.py -m train -d unc -t train -f ckpts/unc
To test the model with Dense CRF, run
python main_convlstm_p543.py -m test -d unc -t testA -f ckpts/unc -i 700000 -c
If you use this code, please consider citing
@inproceedings{li2018referring,
title={Referring Image Segmentation via Recurrent Refinement Networks},
author={Li, Ruiyu and Li, Kaican and Kuo, Yi-Chun and Shu, Michelle and Qi, Xiaojuan and Shen,
Xiaoyong and Jia, Jiaya},
booktitle={CVPR},
year={2018}
}