Skip to content

Latest commit

 

History

History

AGORA

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
README.md
README.md
 
 
agora2coco_smpl.py
agora2coco_smpl.py
 
 
agora2coco_smplx.py
agora2coco_smplx.py
 
 
crop_and_resize_4k_images.py
crop_and_resize_4k_images.py
 
 

README.md

AGORA dataset parsing process

  • All files can be downloaded from here.
  • If you want only one of SMPL and SMPLX data, you can ignore the other one.
  • This repo requires only SMPLX data, while Pose2Pose branch requires only SMPL data.

Make annotation files

  • This code will dump SMPL/SMPLX parameters in camera-centered coordinate system and camera parameters at smpl_params_cam/smplx_params_cam and cam_params folders, respectively, in $DATASET_PATH. Also, it will generate AGORA_train_SMPL.json, AGORA_validation_SMPL.json, AGORA_train_SMPLX.json, and AGORA_validation_SMPLX.json in $DATASET_PATH.
  • For the SMPL data, 1) download and unzip smpl_gt.zip, train_SMPL.zip, and validation_SMPL.zip and 2) run python agora2coco_smpl.py --dataset_path $DATASET_PATH.
  • For the SMPLX data, 1) download and unzip smplx_gt.zip, train_SMPLX.zip and validation_SMPLX.zip and 2) run python agora2coco_smplx.py --dataset_path $DATASET_PATH.
  • $DATASET_PATH denotes AGORA dataset path.

Preparing 1280x720 image files

  • This code will prepare 1280x720 image files.
  • Download and unzip 1280x720 image files.
  • Then, make images_1280x720 folder in AGORA dataset path.
  • For the $i$th zip file of training set, make train_$i$ folder and move all image files to that folder. For example, make train_0 folder at AGORA dataset path and move all image files from train_images_1280x720_0.zip to that folder.
  • For the images of validation and test sets, make validation and test folders and move all images files to corresponding folders.

Preparing 3840x2160 image files

  • This code will prepare 3840x2160 image files.
  • Do the same process of 1280x720 image files
  • As the image resolution is too high, you need to crop and resize humans to prevent the dataloader from being stuck.
  • To this end, run python crop_and_resize_4k_images.py --dataset_path $DATASET_PATH --out_height 512 --out_width 384. $DATASET_PATH denotes AGORA dataset path.

Download AGORA_test_bbox.json

  • Download human detection results on test set from here.
  • The human detection results are from YOLO v5.

Final directory

${DATASET_PATH}
|-- AGORA_train_SMPL.json
|-- AGORA_validation_SMPL.json
|-- AGORA_train_SMPLX.json
|-- AGORA_validation_SMPLX.json
|-- AGORA_test_bbox.json
|-- smpl_params_cam
|   |-- train_0
|   |-- train_1
|   |-- train_2
|   |-- train_3
|   |-- train_4
|   |-- train_5
|   |-- train_6
|   |-- train_7
|   |-- train_8
|   |-- train_9
|   |-- validation
|-- smplx_params_cam
|   |-- train_0
|   |-- train_1
|   |-- train_2
|   |-- train_3
|   |-- train_4
|   |-- train_5
|   |-- train_6
|   |-- train_7
|   |-- train_8
|   |-- train_9
|   |-- validation
|-- cam_params
|   |-- train_0
|   |-- train_1
|   |-- train_2
|   |-- train_3
|   |-- train_4
|   |-- train_5
|   |-- train_6
|   |-- train_7
|   |-- train_8
|   |-- train_9
|   |-- validation
|-- images_1280x720
|   |-- train_0
|   |-- train_1
|   |-- train_2
|   |-- train_3
|   |-- train_4
|   |-- train_5
|   |-- train_6
|   |-- train_7
|   |-- train_8
|   |-- train_9
|   |-- validation
|   |-- test
|-- images_3840x2160
|   |-- train_0
|   |-- train_0_crop
|   |-- train_1
|   |-- train_1_crop
|   |-- train_2
|   |-- train_2_crop
|   |-- train_3
|   |-- train_3_crop
|   |-- train_4
|   |-- train_4_crop
|   |-- train_5
|   |-- train_5_crop
|   |-- train_6
|   |-- train_6_crop
|   |-- train_7
|   |-- train_7_crop
|   |-- train_8
|   |-- train_8_crop
|   |-- train_9
|   |-- train_9_crop
|   |-- validation
|   |-- validation_crop
|   |-- test
|   |-- test_crop