garment-extraction.md

Garment Extraction from Fashion Images (course project from 3DV of ETH Zürich)

Please check our Report for more details.

Given segmentation json along with its corresponding image, it is possible to extract only garment geometries from the ICON's full reconstruction.

Format of segmentation json:

  {
  "item1": {
    "segmentation": [
      [
        x1,y1,x2,y2,.......
      ],
      [
        .....
      ]
    ],
    "category_id": 0,
    "category_name": "{clothing description}"
  },
  "item2": {
    "segmentation": [
      [
        x1,y1,x2,y2,.......
      ],
      [
        .....
      ]
    ],
    "category_id": 0,
    "category_name": "{clothing description}"
  }
}

Where "x1, y1, ..." are the coordinates of the segmentation boundaries. "category_id" identifies the type of clothing (refer to DeepFashion2). The segmentation and image files must have the same name. Please check a few samples under the examples/segmentation.

To run ICON with segmentation you need to add -seg_dir {path_to_segmentation_folder}, all the output results are loccated under {out_dir}/{method}/clothes.

# Inference with segmentation
python -m apps.infer -cfg ./configs/icon-filter.yaml -gpu 0 -in_dir ./examples/segmentation -out_dir ./results -seg_dir ./examples/segmentation

# output under results/icon-filter/clothes

Citation

@techreport{Gudmundsson2022garment,
  title       = "3D Garments Crawler from large-scale fashion images",
  author      = "Gudmundsson, Daniel and Barrau-Joyeaux, Marion and Collette, Arthur and Kjaer, Amalie and Peng, Songyou and Xiu, Yuliang",
  institution = "ETH Zürich",
  year        = 2022,
  month       = jun
}

Contact

For questions, please contact with Daniel Gudmundsson, Marion Barrau-Joyeaux, Arthur Collette, and Amalie Kjaer.