# refseg_rrn

**Repository Path**: cyanlaser/refseg_rrn

## Basic Information

- **Project Name**: refseg_rrn
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2021-11-18
- **Last Updated**: 2021-11-23

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# (reproduce) Referring Image Segmentation via Recurrent Refinement Networks

This codebase modified from [RRN official Repo](https://github.com/liruiyu/referseg_rrn).

## Setup

This code derives from [TF-phrasecut-public](https://github.com/chenxi116/TF-phrasecut-public).
Please follow its `Setup` and `Data Preparation` sections except that the DeepLab-ResNet backbone
comes from [tensorflow-deeplab-resnet](https://github.com/DrSleep/tensorflow-deeplab-resnet). The
pre-trained DeepLab-ResNet model can be downloaded from
[here](https://github.com/DrSleep/tensorflow-deeplab-resnet#caffe-to-tensorflow-conversion).

### Install Runtime Packages

1. Install Runtime Environment:
  1. conda create -n tf1.2 python=3.6;
  2. conda install cudatoolkit=8.0;
  3. conda install cudnn=5.1;
  4. pip install tensorflow-gpu==1.2.1
2. pip install -r requirements.txt;

### Prepare backbone pretrained weights

1. visit official setup of [tensorflow-deeplab-resnet](https://github.com/DrSleep/tensorflow-deeplab-resnet#caffe-to-tensorflow-conversion);
2. open [google drive](https://drive.google.com/drive/folders/0B_rootXHuswsZ0E4Mjh1ZU5xZVU?resourcekey=0-9Ui2e1br1d6jymsI6UdGUQ) ;
3. Download `deeplab_resnet_init.ckpt` and put it into `data/weights/`;

### Install REFER API

1. cd `external/refer`;
2. Download `refclef`, `refcoco`, `refcoco+`, `refcocog` from:
    1. https://bvisionweb1.cs.unc.edu/licheng/referit/data/refclef.zip
    2. https://bvisionweb1.cs.unc.edu/licheng/referit/data/refcoco.zip
    3. https://bvisionweb1.cs.unc.edu/licheng/referit/data/refcoco+.zip
    4. https://bvisionweb1.cs.unc.edu/licheng/referit/data/refcocog.zip
3. make;

## Usage
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
```

## Thanks

1. official repo: https://github.com/liruiyu/referseg_rrn