sy000/ChatGarmentDataset

ChatGarment Dataset

It is only a temporary version. We will re-organize it later.

Introduction

We generated four versions of the dataset for ChatGarment. Based on quality assessment, we use only v1, v2 and v4 for training in the final ChatGarment model.

Data format for v1 & v2

We utilize the GarmentCodeData v1 dataset as the base garment dataset for garment sampling.

The labels are given in garment_restpose_data_v1.json (v1) and garment_restpose_data_v1.json (v2).

Here is the structure of the json files:

[{
    'upper_garment': ... // Directory of upper garments in GarmentCodeData v1 dataset
    'lower_garment': ... // Directory of lower garments in GarmentCodeData v1 dataset
    'whole_garment': ... // Directory of wholebody garments in GarmentCodeData v1 dataset
    'text_label': ... // Text descriptions for the outfit
    'img_name': ... // Directory of outfit images in garments_imgs_v1 or garments_imgs_v2
    'img_restpose_name': // Directory of A-pose outfit images in garments_imgs_v1 or garments_imgs_v2
},
...
]

Dataset v1 contains only A-pose garments, and the text labels are extracted directly from dataset_properties_default_body.yaml in GarmentCodeData v1 dataset.

Dataset v2 contains garments in different poses. The text labels are generated by GPT-4o.

Data format for v4

We resample the data using customized rules.

The labels are given in garment_restpose_data_v4.json.

Here is the structure of the json files:

[{
    'upper_garment': ... // Directory of upper garments in garments_imgs_v4
    'lower_garment': ... // Directory of lower garments in garments_imgs_v4
    'whole_garment': ... // Directory of wholebody garments in garments_imgs_v4
    'text_label_list': ... // A list of slightly different text descriptions for the outfit
    'img_name': ... // Directory of outfit images in garments_imgs_v4
},
...
]

Dataset v4 contains garments in different poses. The text labels are generated by GPT-4o.

Citation

@article{bian2024chatgarment,
  title={ChatGarment: Garment Estimation, Generation and Editing via Large Language Models},
  author={Bian, Siyuan and Xu, Chenghao and Xiu, Yuliang and Grigorev, Artur and Liu, Zhen and Lu, Cewu and Black, Michael J and Feng, Yao},
  journal={arXiv preprint arXiv:2412.17811},
  year={2024}
}