Skip to content

Latest commit

 

History

History
62 lines (47 loc) · 2.99 KB

DATASETS.md

File metadata and controls

62 lines (47 loc) · 2.99 KB
Raw

Dataset preparation

Configuration

Please create a config file called semantic_segmentation.cfg that specifies the paths to the datasets. Note that they are optional; you don't need to download and provide all datasets, only the ones you intend to use. Replace the paths below with something that works for you:

[paths]
camvid=/datasets/camvid/CamVidData.zip
cityscapes=/datasets/cityscapes\cityscapes_segmentation.zip
isic2017=/datasets/isic2017/isic2017_segmentation_248x248.zip
pascal_voc=/datasets/pascal_voc2012/VOCdevkit/VOC2012

Note that the CamVid, Cityscapes and ISIC 2017 datasets must be converted to a ZIP-based format prior to use. You must run the provided conversion utilities to create these ZIP files.

Pascal VOC 2012

  1. Download the Pascal VOC 2012 dataset (use the 'training/validation data' link).
  2. You will also want the augmented labels (download here) so you can use the augmented Pascal dataset (used in Mittal et al. and Hung et al.)
  3. Decompress the main main dataset file VOCtrainval_11-May-2012.tar
  4. Unzip SegmentationClassAug.zip within the VOCdevkit/VOC2012 directory that was created by unpacking the main dataset.
  5. Edit the semantic_segmentation.cfg configuration file and provide a path for the pascal_voc setting.
  6. Now run: python download_pascal_aug_names.py to download some index files

The specific split used in Mittal et al. can be found in data/splits/pascal_aug/split_0.pkl. This file was taken as-is from their repo.

Cityscapes

  1. Sign up for a cityscapes account at https://www.cityscapes-dataset.com/
  2. Download the input images file leftImg8bit_trainvaltest.zip
  3. Download the ground truth file gtFine_trainvaltest.zip.
  4. Edit the semantic_segmentation.cfg configuration file and point the cityscapes to a place where you want the converted Cityscapes ZIP to live
  5. Run: python convert_cityscapes.py /path/to/leftImg8bit_trainvaltest.zip /path/to/gtFine_trainvaltest.zip

The conversion process will downsample all images by a factor of 2 as in Mittal et al. and Hung et al..

ISIC 2017

  1. Download the ISIC 2017 zip files: ISIC-2017_Training_Data.zip, ISIC-2017_Training_Part1_GroundTruth.zip, ISIC-2017_Validation_Data.zip and ISIC-2017_Validation_Part1_GroundTruth.zip to a directory called e.g. /path/to/isic_zips_directory.
  2. Run: python convert_isic.py /path/to/isic_zips_directory

The conversion process will scale all images to a default size of 248x248. (Use the --out_size=<height>,<width> when running convert_isic to change this).