Paper

Pytorch implementation of our paper On exploring weakly supervised domain adaptation strategies for semantic segmentation using synthetic data.

Preparation

Pre-requisites

• Python 3.7
• Pytorch >= 0.4.1
• CUDA 9.0 or higher

Installation

0. Clone the repo:

$ git clone https://github.com/RobertoAlcoverCouso/OEWSDA
$ cd OEWSDA

1. Create a conda environment:

$ conda create -n OEWSDA python=3.7
$ conda activate OEWSDA

2. Install OpenCV and pytorch if you don't already have it:

$ conda install pytorch=0.4.1 cuda90 torchvision -c pytorch
$ conda install -c menpo opencv

3. Install this repository and the dependencies using pip <root_dir> stands for ./ if you follow the instructions:

$ pip install -e <root_dir> 

Datasets

By default, the datasets are put in <root_dir>/../data. An alternative option is to explicitlly specify the parameters in the cfg file.

• GTA5: Please follow the instructions here to download images and semantic segmentation annotations. The GTA5 dataset directory should have this basic structure:

<root_dir>/../data/GTA5/                               % GTA dataset root
<root_dir>/../data/GTA5/images/                        % GTA images
<root_dir>/../data/GTA5/labels/                        % Semantic segmentation labels
...

• Cityscapes: Please follow the instructions in Cityscapes to download the images and validation ground-truths. The Cityscapes dataset directory should have this basic structure:

<root_dir>/../data/Cityscapes/                         % Cityscapes dataset root
<root_dir>/../data/Cityscapes/leftImg8bit              % Cityscapes images
<root_dir>/../data/Cityscapes/leftImg8bit/val
<root_dir>/../data/Cityscapes/gtFine                   % Semantic segmentation labels
<root_dir>/../data/Cityscapes/gtFine/val
...

Prepare the datasets

For each dataset analized, there is a <dataset_name>_utils.py file in the <root_dir>/dataset folder. This file will transform the semantic labels to the Cityscapes train labels to be employed. For example for the Cityscapes dataset:

$ python  dataset/cityscapes_utils.py

This should have created 3 csvs for each of the subsets: "trainCS.csv", "valCS.csv" and "testCS.csv"

Train

• Model weights: If you intend on using deeplab please download the imagenet-pretrained backbone weights.

Additional weights can be found in the publication page

• Train: Modify the yalm file corresponding to the experiment you want to run "<experiment_name>.yalm" to include the datasets you want to train with, the proportion to use in the range of 0-1 as follows and the model you want to use as follows:

architecture: <architecture_name>
train_set:
    <dataset_1>: <proportion_of_dataset_1>
    <dataset_2>: <proportion_of_dataset_2>
    ...

<architecture_name> is expected to be one of the following: "deeplabv3","FCN" or "psp" To train run the command line:

python main.py --config config/<experiment_name>.yalm

Note that for fine_tuning a restore file is expected in the "restore_file" argument.

Validate

run the command line:

python main.py -r <model_to_validate> --config config/validate.yalm

Model weights: