CosmicManHQ-1.0 Dataset

Name: CosmicManHQ-1.0
Creator: CosmicMan
License: https://choosealicense.com/licenses/cc-by-4.0/

Overview

The CosmicMan-HQ 1.0 dataset is a large-scale, high-quality human image dataset with a mean resolution of 1488 × 1255 pixels. It is equipped with rich text annotations and is designed to support a variety of human-centric content generation tasks. The number of images released in current dataset is 5,463,500 and all these images are sourced from LAION-5B.

Data Structure

Now we release the human parsing (in folder './LAION-5B-Pasing/') and annotations (in folder './label/') for CosmicManHQ-1.0. The human parsing paths "segm_path" are added in the annotations.

Each annotation adheres to the subsequent Parquet format specifications, including column names and corresponding content examples:

    {
        "image_name": "3348418094672.jpg",
        "height": "1028",
        "width": "822",
        "person_type": "full_body",
        "similarity": "0.29458436369895935",
        "url": "https://cdn0.matrimonio.com.co/img_g/articulos-colombia/2014-06-26-vestidos-corte-imperio/111-umbria.jpg",
        "segm_path": "LAION-5B-Parsing/laion1B-nolang/00044/3348418094672.png",
        "path": "LAION-5B/laion1B-nolang/00044/3348418094672.jpg",
        "raw_text": "vestido de novia imperio",
        "blip_text": "a woman in a white dress standing in front of a window",
        "aa_text": [
            {
                "gender": "female",
                "country": "caucasian",
                "age": "adult"
            },
            {
                "body shape": "fit"
            },
            ...
        ],
        "aa_ids": [
            [
                11
            ],
            [
                0
            ],
            ...
        ]
    },

"height": the height of image.
"width": the width of image.
"person_type": the type of image, contain ["full_body", "upper_body", "portrait", "headshot", "nearly_full_body", "others"].
"similarity": the CLIP similarity of image and raw_text.
"url": the url of image.
"segm_path": the path of human parsing.
"path": the path of image from Laion-5B dataset.
"raw_text": the description from Laion-5B dataset.
"blip_text": the description from Blip.
"aa_text": we initially established a set of predefined body parts, where each item in the "aa_text" list corresponds to a specific part. The key-value pairs for each item are associated with a particular question. For more detailed information, please refer to the file './details_of_aa_text.json'. It is important to note that in the "aa_text", only the parts with a basic inquiry (e.g., "Does the person wear any tops?") that have an affirmative response ("yes") are retained. And the "wears" key itself is not included within the "aa_text".
"aa_ids": we utilized the SCHP model trained on the ATR dataset to extract results for each body part (as shown in the table below) and then integrated these results with "aa_text" to ascertain whether each part has a corresponding parsing ID and annotation. In other words, if the basic inquiry for body part i receives a "no" response, or if the all potential parsing ID is absent from the parsing results, then both "aa_text[i]" and "aa_ids[i]" are designated as "NULL".

Predefined Body Part	Potential Parsing Part	Potential Parsing ID
face	face	11
body shape	in addition to the background	0 (used to invert to get the foreground)
background	background	0
hair	hair	2
special clothings	upper-clothes,skirt,pants,dress	4,5,6,7
one-piece outfits	upper-clothes,skirt,pants,dress	4,5,6,7
tops	upper-clothes	4
coats	upper-clothes	4
bottoms	skirt,pants	5,6
shoes	left-shoe,right-shoe	9,10
bags	bag	16
hats	hat	1
belts	belt	8
scarf	scarf	17
others	-	-

Download Instructions

Download Human Parsing Images and Annotations

The human parsing images are placed in folder ',.LAION-5B-Pasing' of the main directory ./CosmicManHQ-1.0/ and the annotations are placed in folder './label/' of the main directory ./CosmicManHQ-1.0/. You can obtain these resources by directly downloading this repository. Please refer to this page for download instructions.

Download Images

The images in CosmicManHQ-1.0 Dataset are source from LAION-5B and we will not directly provide these images in this repository. As an alternative，each image download url is provided in the annotations and you can download the images by following way:

Install the 'img2dataset' package provided in this document, which supports specifying a unique save path for each downloaded image.

pip install -e ./img2dataset/

You need to first download the parquet files (in folder './label/'). Next, please using the file './download.py' in the main directory ./CosmicManHQ-1.0/ to download the images. More details refer to the './download.py'.

mkdir -p CosmicManHQ-1.0/ && cd CosmicManHQ-1.0/

python download.py \
    --start_index 0 \
    --end_index 256 \
    --parquet_dir label/ \
    --log_dir download_log/

rm -rf download_log/

Citation Information

@article{li2024cosmicman,
  title={CosmicMan: A Text-to-Image Foundation Model for Humans},
  author={Li, Shikai and Fu, Jianglin and Liu, Kaiyuan and Wang, Wentao and Lin, Kwan-Yee and Wu, Wayne},
  journal={arXiv preprint arXiv:2404.01294},
  year={2024}
}

Reference

Li P, Xu Y, Wei Y, et al. Self-correction for human parsing[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020, 44(6): 3260-3271.
Schuhmann C, Beaumont R, Vencu R, et al. Laion-5b: An open large-scale dataset for training next generation image-text models[J]. Advances in Neural Information Processing Systems, 2022, 35: 25278-25294.
Img2dataset: https://github.com/rom1504/img2dataset

Contributors

Shikai Li, Jianglin Fu, Kaiyuan Liu, Wentao Wang, Kwan-Yee Lin, Wayne Wu

Datasets:

cosmicman
/

CosmicManHQ-1.0

You need to agree to share your contact information to access this dataset