AIST++ API
This repo contains starter code for using the AIST++ dataset. To download the dataset or explore details of this dataset, please go to our dataset website.
Installation
The code has been tested on python>=3.7. You can install the dependencies and this repo by:
pip install -r requirements.txt
python setup.py install
You also need to make sure ffmpeg is installed on your machine, if you would like to visualize the annotations using this api.
How to use
We provide demo code for loading and visualizing AIST++ annotations. Note AIST++ annotations and videos, as well as the SMPL model (for SMPL visualization only) are required to run the demo code.
The directory structure of the data is expected to be:
├── motions/
├── keypoints2d/
├── keypoints3d/
├── splits/
├── cameras/
└── ignore_list.txt
└── *.mp4
├── SMPL_MALE.pkl
└── SMPL_FEMALE.pkl
Visualize 2D keypoints annotation
The command below will plot 2D keypoints onto the raw video and save it to the directory ./visualization/.
python demos/run_vis.py \
--anno_dir <ANNOTATIONS_DIR> \
--video_dir <VIDEO_DIR> \
--save_dir ./visualization/ \
--video_name gWA_sFM_c01_d27_mWA2_ch21 \
--mode 2D
Visualize 3D keypoints annotation
The command below will project 3D keypoints onto the raw video using camera parameters, and save it to the directory ./visualization/.
python demos/run_vis.py \
--anno_dir <ANNOTATIONS_DIR> \
--video_dir <VIDEO_DIR> \
--save_dir ./visualization/ \
--video_name gWA_sFM_c01_d27_mWA2_ch21 \
--mode 3D
Visualize the SMPL joints annotation
The command below will first calculate the SMPL joint locations from our motion annotations (joint rotations and root trajectories), then project them onto the raw video and plot. The result will be saved into the directory ./visualization/.
python demos/run_vis.py \
--anno_dir <ANNOTATIONS_DIR> \
--video_dir <VIDEO_DIR> \
--smpl_dir <SMPL_DIR> \
--save_dir ./visualization/ \
--video_name gWA_sFM_c01_d27_mWA2_ch21 \
--mode SMPL
Multi-view 3D keypoints and motion reconstruction
This repo also provides code we used for constructing this dataset from the multi-view AIST Dance Video Database. The construction pipeline starts with frame-by-frame 2D keypoint detection and manual camera estimation. Then triangulation and bundle adjustment are applied to optimize the camera parameters as well as the 3D keypoints. Finally we sequentially fit the SMPL model to 3D keypoints to get a motion sequence represented using joint angles and a root trajectory. The following figure shows our pipeline overview.
AIST++ construction pipeline overview.
The annotations in AIST++ are in COCO-format for 2D & 3D keypoints, and SMPL-format for human motion annotations. It is designed to serve general research purposes. However, in some cases you might need the data in different format (e.g., Openpose / Alphapose keypoints format, or STAR human motion format). With the code we provide, it should be easy to construct your own version of AIST++, with your own keypoint detector or human model definition.
Step 1. Assume you have your own 2D keypoint detection results stored in .pkl format that we support. The code we used at this step is as follows but you might need to modify the script run_preprocessing.py in order to be compatible with your own data.
python processing/run_preprocessing.py \
--keypoints_dir <KEYPOINTS_DIR> \
--save_dir <ANNOTATIONS_DIR>/keypoints2d/
Step 2. Then you can estimate the camera parameters using your 2D keypoints. This step is optional as you can still use our camera parameter estimates which are quite accurate. At this step, you will need the
# If you would like to estimate your own camera parameters:
python processing/run_estimate_camera.py \
--anno_dir <ANNOTATIONS_DIR> \
--save_dir <ANNOTATIONS_DIR>/cameras/
# Or you can skip this step by just using our camera parameter estimates.
Step 3. Next step is to perform 3D keypoints reconstruction from multi-view 2D keypoints and camera parameters. You can just run:
python processing/run_estimate_keypoints.py \
--anno_dir <ANNOTATIONS_DIR> \
--save_dir <ANNOTATIONS_DIR>/keypoints3d/
Step 4. Finally we can estimate SMPL-format human motion data by fitting the 3D keypoints to the SMPL model. If you would like to use another human model such as STAR, you will need to do some modifications in the script run_estimate_smpl.py. The following command runs SMPL fitting.
python processing/run_estimate_smpl.py \
--anno_dir <ANNOTATIONS_DIR> \
--smpl_dir <SMPL_DIR> \
--save_dir <ANNOTATIONS_DIR>/motions/
Note that this step will take several days to process the entire dataset if your machine has only one GPU. In practise, we run this step on a cluster, but are only able to provide the single-threaded version.
MISC.
- COCO-format keypoint definition:
[
"nose",
"left_eye", "right_eye", "left_ear", "right_ear", "left_shoulder","right_shoulder",
"left_elbow", "right_elbow", "left_wrist", "right_wrist", "left_hip", "right_hip",
"left_knee", "right_knee", "left_ankle", "right_ankle"
]
- SMPL-format body joint definition:
[
"root",
"left_hip", "left_knee", "left_foot", "left_toe",
"right_hip", "right_knee", "right_foot", "right_toe",
"waist", "spine", "chest", "neck", "head",
"left_in_shoulder", "left_shoulder", "left_elbow", "left_wrist",
"right_in_shoulder", "right_shoulder", "right_elbow", "right_wrist"
]
However, when the camera translations of the annotations are displayed in 3d space, they appear as follows compared to other keypoints.
You can see that the height does not match with the human keypoints, and it is not surrounded by 360 degrees and is clustered in a specific area.
342 Dec 27, 2022
1 Jan 10, 2022
7 Oct 30, 2022
5 Apr 9, 2022
104 Dec 29, 2022
44 Nov 15, 2022
1 Nov 21, 2021
2 Jan 12, 2022
2 Jan 15, 2022
1 Jan 24, 2022
19 Apr 24, 2022
207 Dec 30, 2022
28 May 23, 2022
2 Oct 15, 2021
3 Nov 2, 2022
0 Aug 20, 2022
82 Dec 26, 2022
2 Jun 9, 2022
2 Feb 2, 2022