픽토그램 포즈 기반 사람 이미지 생성 모델

Installation

Build Environment

We recommend a python version =3.9.12 and cuda version =11.6.

pip install -r requirements.txt

Download Weights

1. Download trained weights from Google Drive

2. Then, these weights should be organized as follows:

   model
   |-- omnipose
   |   |-- checkpoint_ppe.pth
   |   |-- checkpoint_hpe.pth
   |   `...
   |-- pg2
   |   |-- checkpoints
   |   |   |-- PG2-1
   |   |   |   |-- events.out.tfevents.1566133862.wangbudui.39385.0
   |   |   |   `-- G1.pth
   |   |   `-- PG2-2
   |   |       |-- events.out.tfevents.1566139204.wangbudui.40174.0
   |   |       `-- G2.pth
   |   `...
   `...
   

Inference

Method 1) With Flask Server

1. Prepare .env file to ./backend

   ACCESS_KEY={s3_access_key}
   SECRET_KEY={s3_secret_key}
   REGION_KEY={s3_region}
   BUCKET_NAME={s3_bucket_name}

2. run ./backend/app.py

3. request api to flask server

   • Method: POST
   • URI: '/pgpg'

Method 2) Without Flask Server

# 1. Declare variables
generation_id = {any_integer}
condition_image_url = {your_condition_image_url}
target_image_url = {your_target_image_url}

# 2. HPE + PPE
omnipose.run_infer.run_infer(generation_id, condition_image_url, target_image_url)

# 3. mpii2coco + generate annotation, pairs
pe2pg.Pe2Pg(generation_id).save_anno_csv()

# 4. generate posemap & mask
pg2.tool.generate_pose_map_add_mask.main("market", "inference", generation_id)

# 5. PG2
pg2.run.main(generation_id)

What We Modified

Files we modified from the existing implementation

• OmniPose
• PG2

Members

강병욱, 김태인, 석예은, 윤지현, 이진주, 한상우, 홍정현