First step, obtain your login credentials.

Here we give Five different datasets example including Mapillary Cityscapes, Camvid, KITTI and BDD.

To be note that we only use the pretrained WiderResNet model on Mapillary for fair comparison on Cityscapes.

Mapillary Vistas Dataset

First of all, please request the research edition dataset from here. The downloaded file is named as mapillary-vistas-dataset_public_v1.1.zip.

Then simply unzip the file by

unzip mapillary-vistas-dataset_public_v1.1.zip

The folder structure will look like:

Mapillary
├── config.json
├── demo.py
├── Mapillary Vistas Research Edition License.pdf
├── README
├── requirements.txt
├── training
│   ├── images
│   ├── instances
│   ├── labels
│   ├── panoptic
├── validation
│   ├── images
│   ├── instances
│   ├── labels
│   ├── panoptic
├── testing
│   ├── images
│   ├── instances
│   ├── labels
│   ├── panoptic

Note that, the instances, labels and panoptic folders inside testing are empty.

Cityscapes Dataset

Download Dataset

First of all, please request the dataset from here. You need multiple files. Both Coarse data and Fine data are used. Note that we do not use coarse data for training but we use it for uniform sampling.

- leftImg8bit_trainvaltest.zip
- gtFine_trainvaltest.zip
- leftImg8bit_trainextra.zip
- gtCoarse.zip
- leftImg8bit_sequence.zip    # This file is very large, 324G. You only need it if you want to run sdc_aug experiments. 

If you prefer to use command lines (e.g., wget) to download the dataset,

# First step, obtain your login credentials.
Please register an account at https://www.cityscapes-dataset.com/login/.

# Second step, log into cityscapes system, suppose you already have a USERNAME and a PASSWORD.
wget --keep-session-cookies --save-cookies=cookies.txt --post-data 'username=USERNAME&password=PASSWORD&submit=Login' https://www.cityscapes-dataset.com/login/

# Third step, download the zip files you need.
wget -c -t 0 --load-cookies cookies.txt --content-disposition https://www.cityscapes-dataset.com/file-handling/?packageID=3

# The corresponding packageID is listed below,
1  -> gtFine_trainvaltest.zip (241MB)            md5sum: 4237c19de34c8a376e9ba46b495d6f66
2  -> gtCoarse.zip (1.3GB)                       md5sum: 1c7b95c84b1d36cc59a9194d8e5b989f
3  -> leftImg8bit_trainvaltest.zip (11GB)        md5sum: 0a6e97e94b616a514066c9e2adb0c97f
4  -> leftImg8bit_trainextra.zip (44GB)          md5sum: 9167a331a158ce3e8989e166c95d56d4
14 -> leftImg8bit_sequence.zip (324GB)           md5sum: 4348961b135d856c1777f7f1098f7266

Prepare Folder Structure

Now unzip those files, the desired folder structure will look like,

Cityscapes
├── leftImg8bit_trainvaltest
│   ├── leftImg8bit
│   │   ├── train
│   │   │   ├── aachen
│   │   │   │   ├── aachen_000000_000019_leftImg8bit.png
│   │   │   │   ├── aachen_000001_000019_leftImg8bit.png
│   │   │   │   ├── ...
│   │   │   ├── bochum
│   │   │   ├── ...
│   │   ├── val
│   │   ├── test
├── gtFine_trainvaltest
│   ├── gtFine
│   │   ├── train
│   │   │   ├── aachen
│   │   │   │   ├── aachen_000000_000019_gtFine_color.png
│   │   │   │   ├── aachen_000000_000019_gtFine_instanceIds.png
│   │   │   │   ├── aachen_000000_000019_gtFine_labelIds.png
│   │   │   │   ├── aachen_000000_000019_gtFine_polygons.json
│   │   │   │   ├── ...
│   │   │   ├── bochum
│   │   │   ├── ...
│   │   ├── val
│   │   ├── test
├── leftImg8bit_trainextra
│   ├── leftImg8bit
│   │   ├── train_extra
│   │   │   ├── augsburg
│   │   │   ├── bad-honnef
│   │   │   ├── ...
├── gtCoarse
│   ├── gtCoarse
│   │   ├── train
│   │   ├── train_extra
│   │   ├── val
├── leftImg8bit_sequence
│   ├── train
│   ├── val
│   ├── test

CamVid Dataset

Please download and prepare this dataset according to the tutorial. The desired folder structure will look like,

CamVid
├── train
├── trainannot
├── val
├── valannot
├── test
├── testannot

KITTI Dataset

Please download this dataset at the KITTI Semantic Segmentation benchmark webpage.

Now unzip the file, the desired folder structure will look like,

KITTI
├── training
│   ├── image_2
│   ├── instance
│   ├── semantic
├── test
│   ├── image_2

There is no official training/validation split as the dataset only has 200 training samples. We randomly create three splits at here in order to perform cross-validation.

BDD Dataset

Please download this dataset at BDD webpage

and we use the semantic segmentation parts. Unzip the file, the desired folder structure will look like this.

BDD seg
├── images
│   ├── train
│   ├── val
│   ├── test
├── labels
│   ├── train
│   ├── val

After that, you can either change the config.py or do the soft link according to the default path in config.

For example,

Suppose you store your dataset at ~/username/data/Cityscapes, please update the dataset path in config.py,

__C.DATASET.MAPILLARY_DIR = '~/username/data/Cityscapes'

For example, You can link the data path into current folder.

mkdir data 
cd data
ln -s your_cityscapes_root_data_path cityscapes
ln -s your_camvid_root_data_path camvid