Skip to content

Latest commit

 

History

History
66 lines (60 loc) · 2.84 KB

dataset_prepare.md

File metadata and controls

66 lines (60 loc) · 2.84 KB

Prepare datasets

Before generate variations of exisiting segmentation dataset, you should download original datasets first. Then in the dataset preprocess stage, we select one salient object from each sample to conduct local editing in order to guarantee the quality of edited images. We support experiments on PASCAL VOC 2021, ADE20K, COCO Stuff 164k and Cityscapes.

Pascal VOC

  1. Pascal VOC 2012 could be downloaded from here. Please run dataset preparation scripts from mmsegmentation.
  2. Before dataset prepapre, please re-organize the file structure like this: 
    data
├── pascal_voc
│   ├── images
│   │   │── train
│   │   │── val
│   ├── annotations_mmseg
│   │   ├── train
│   │   └── val
  3. Then please run following command to convert files into proper format. 
    python datasets_prepare/pascal_voc.py --data-root $your_data_path --data-save-root $your_data_save_path

ADE20K

  1. The ADE20K dataset could be download from here.
  2. Then please run following command to convert files into proper format. 
    python datasets_prepare/ade20k.py --data-root $your_data_path --data-save-root $your_data_save_path

COCO Stuff 164k

  1. For downloading data and converting annotations, please refer to guidance from mmsegmentation
  2. After prepare the dataset, the files should be 
    data
├── coco_stuff164k
│   ├── images
│   │   │── train2017
│   │   │── val2017
│   ├── annotations
│   │   ├── train2017
│   │   ├── val2017
│   │   ├── train2017.json 
│   │   └── val2017.json
  3. Then please run following command to convert files into proper format. 
    python datasets_prepare/coco_stuff164k.py --data-root $your_data_path --data-save-root $your_data_save_path

After dataset preprocess, the files of each dataset should be like this:

data
├── datasets_name
│   ├── images
│   │   │── train
│   │   │── val
│   ├── annotations # semantic segmentation annotations in mmseg format
│   │   ├── train
│   │   └── val
│   ├── masks_object # mask within [0, 255]
│   │   ├── train
│   │   └── val
│   ├── meta_train.json # includes categories info of selected objects
│   ├── meta_val.json