简体中文 | English

Prepare Public Dataset

For public dataset, you only need to download and save it in specific directory, and then use PaddleSeg to train model.

Save Directory

PaddleSeg defines the dataset path in config files according the following directory.

It is recommended to download and save public dataset in PaddleSeg/data.

PaddleSeg
├── data
│   ├── cityscapes
│   │   ├── leftImg8bit
│   │   │   ├── train
│   │   │   ├── val
│   │   ├── gtFine
│   │   │   ├── train
│   │   │   ├── val
│   ├── ADEChallengeData2016
│   │   │── annotations
│   │   │   ├── training
│   │   │   ├── validation
│   │   │── images
│   │   │   ├── training
│   │   │   ├── validation
│   ├── VOCdevkit
│   │   ├── VOC2012
│   │   │   ├── JPEGImages
│   │   │   ├── SegmentationClass
│   │   │   ├── SegmentationClassAug
│   │   │   ├── ImageSets
│   │   │   │   ├── Segmentation

Download Public Dataset

CityScapes Dataset

Cityscapes is a dataset of semantically understood images of urban street scenes. It mainly contains street scenes from 50 different cities, with 5000 (2048 x 1024) high quality pixel-level annotated images of urban driving scenes. It contains 19 categories. There are 2975 training sets, 500 validation sets and 1525 test sets.

Due to restrictions, please visit CityScapes websiteto download dataset. We recommend that you store dataset in PaddleSeg/data for full compatibility with our config files. Please organize the dataset into the following structure after downloading:

    cityscapes
    |
    |--leftImg8bit
    |  |--train
    |  |--val
    |  |--test
    |
    |--gtFine
    |  |--train
    |  |--val
    |  |--test

Run the following command to convert labels:

pip install cityscapesscripts
python tools/data/convert_cityscapes.py --cityscapes_path data/cityscapes --num_workers 8

where cityscapes_path should be adjusted according to the actual dataset path. num_workers determines the number of processes to be started. The value can be adjusted as required.

Pascal VOC 2012 dataset

Pascal VOC 2012 is mainly object segmentation, including 20 categories and background classes, including 1464 training sets and 1449 validation sets. Generally, we will use SBD(Semantic Boundaries Dataset) to expand the dataset. Theer are 10582 training sets after expanding.

Run the following commands to download the SBD dataset and use it to expand:

cd PaddleSeg
python tools/data/voc_augment.py --voc_path data/VOCdevkit --num_workers 8

where voc_pathshould be adjusted according to the actual dataset path.

ADE20K Dataset

ADE20K published by MIT that can be used for a variety of tasks such as scene perception, segmentation, and multi-object recognition. It covers 150 semantic categories, including 20210 training sets and 2000 validation sets.

Coco Stuff Dataset

Coco Stuff is a pixel-level semantically segmented dataset based on Coco datasets. It covers 172 catefories, including 80 'thing' classes, 91 'stuff' classes amd one 'unlabeled' classes. 'unlabeled' is ignored and the index is set to 255 which has not contribution to loss. The training version is therefore provided in 171 categories. There are 118k training sets, 5k validation sets.

Before using Coco Stuff dataset， please go to COCO-Stuff website to download dataset or download coco2017 training sets with origin images, coco2017 validation sets with origin images and annotations images We recommend that you store dataset in PaddleSeg/data for full compatibility with our config files. Please organize the dataset into the following structure after downloading:

    cocostuff
    |
    |--images
    |  |--train2017
    |  |--val2017
    |
    |--annotations
    |  |--train2017
    |  |--val2017

Run the following command to convert labels:

python tools/data/convert_cocostuff.py --annotation_path /PATH/TO/ANNOTATIONS --save_path /PATH/TO/CONVERT_ANNOTATIONS

where annotation_path should be filled according to the cocostuff/annotations actual path. save_path determines the location of the converted label.

Where, the labels of the labeled images are taken in sequence from 0, 1, ... and cannot be separated. If there are pixels that need to be ignored, they should be labeled to 255.

Pascal Context Dataset

Pascal Context is a pixel-level semantically segmented dataset based on the Pascal VOC 2010 dataset with additional annotations. The conversion script we provide supports 60 categories, with index 0 being the background category. There are 4996 training sets and 5104 verification sets in this dataset.

Before using Pascal Context dataset， Please download VOC2010 firstly，then go to Pascal-Context home pageto download dataset and annotations We recommend that you store dataset in PaddleSeg/data for full compatibility with our config files. Please organize the dataset into the following structure after downloading:

    VOC2010
    |
    |--Annotations
    |
    |--ImageSets
    |
    |--SegmentationClass
    |  
    |--JPEGImages
    |
    |--SegmentationObject
    |
    |--trainval_merged.json

Run the following command to convert labels:

python tools/data/convert_voc2010.py --voc_path /PATH/TO/VOC ----annotation_path /PATH/TO/JSON

where voc_path should be filled according to the voc2010 actual path. annotation_path is the trainval_merged.json saved path.

Where, the labels of the labeled images are taken in sequence from 0, 1, 2, ... and cannot be separated. If there are pixels that need to be ignored, they should be labeled to 255 (default ignored value). When using Pascal Context dataset, Detail need to be installed.