This is a modified repository from dataset-CFPD. Please refer to the original repository for more details.

This dataset is used in the paper, (S. Liu, J. Feng, C. Domokos, H. Xu, J. Huang, Z. Hu, & S. Yan. 2014) CFPD | Fashion parsing with weak color-category labels.

• 2,682 images
• 600 x 400 (height, width)
• pixel-level annotated (segmentation map)
  • 23 categories
  • 13 colors
• make_label.py makes the followings from fashon_parsing_data.mat.
  • label/bbox.json: bounding box (for object detection not semantic segmentation).
  • label/categories.tsv
    • category_id
    • category
• label/main_categories.tsv is selected from categories.tsv for object detection.

1. install the requirements
pip install -r requirements.txt

2. download zip file (name) from author's GoogleDrive
download.sh

3. rename "data" file manually to "data.zip" or run the command below:
rename data data.zip

4. unzip data.zip manually or follow instructions from here (http://stahlworks.com/dev/index.php?tool=zipunzip) to unzip from cmd

5. make label/categories.tsv, label/bbox.json from fashon_parsing_data.mat
python make_label.py

6. Convert .mat to hdf5(.h5), run convert_tmm_to_hdf5.m with matlab

7. Convert .hdf5 files into images with annotations and lists
python export.py

• #refs#
  • 0, 0A~zz: 2,719 groups (each record may be a image info)
    • category_label
      • np.array, float64(actually int), (1, 425)
      • map: super-pix id (1425) -> fine category id (1117)
    • color_label
      • np.array, float64(actually int), (1,425)
      • map: super-pix id (1425) -> fine color id (160)
    • ?img_name
      • np.array, uint16, (2~9 etc, 1)
      • range: 59~108 etc
      • ???
    • segmentation
      • np.array, float32, (400, 600)
      • map: img pix loc (w, h) -> super-pix id (1~425)
      • (height, width) (transposing this) is the right orientation.
  • b~x: 23 datasets (each record means category id)
    • np.array, uint16, (1, #fine_category)
    • range: 1~117
    • fine categories's ids under the cateogry.
  • y,z,A~K: 13 datasets (each record means color id)
    • np.array, uint16, (1, #fine_color)
    • range: 1~60
    • fine color's ids under the color.
• all_category_name
  • np.array, h5py.h5r.Reference, (1, 23)
  • Each reference correspods the above cateogory keys under #refs#.
• all_color_name
  • np.array, h5py.h5r.Reference, (1, 13)
  • Each reference correspods the above color keys under #refs#.

• Mentioned in this paper's Figure 7, (P. Tangseng, Z. Wu, & K. Yamaguchi. 2017) Looking at Outfit to Parse Clothing..
• This repository is the code of the above paper. This code deals with CFPD.