Here, we show EFT fitting procedure by using COCO 2014 as example. You may use other dataset similarly.
- Download extra data from SPIN:
sh scripts/download_spin_data.sh
You can see the downloaded files in "./extradata/spin/"
- Download preprocessed pose_regressor models
sh scripts/download_model_zoo.sh
- Set coco annotation and image files under ./data_sets/coco
.
├── ...
├── data_sets
│ └── coco
│ ├── annotations # the folder with .json files. E.g., person_keypoints_train2014.json
│ └── train2014 # the folder with image files
└── ...
Note that you may use symbolic links to setup the folder structure
- This process converts the raw 2D annotation data into a unified format (in npz)
- Run the following
python -m eft.db_processing.coco
- The output is saved in "./preprocessed_db/coco_2014_train_12kp.npz"
- Run the following
python -m demo.eftFitting
-
You will see a GUI window to see EFT process.
-
Use mouse control to see the 3D in other views (see below about the key information)
-
Toggle "C" in 3D Visualizer window to visualize 3D in image cooridnate
-
Press "q" in 3D Visualizer window to move to the next example
-
Other key information:
- mouse left + move: view change
- mouse right + move: zoom in/out
- shift + mouse left + move: pan
- 'C': toggle between 3D view and image view
- 'q': go to the next sample
- 'w': toggle between solid mesh and wire-frame mesh
- 'j': on/off for 3d skeleton
- 'm': on/off for 3d mesh
- 'f': on/off for floor
- The EFT output is saved in "./eft_out" as PKL format.
- You can disable "--bDebug_visEFT" in eftFitting.py if you only want to get EFT outputs without visualization.
CC-BY-NC 4.0. See the LICENSE file.