EasyMocap: SMPL fitting from multi-camera and internet video, with a dataset gated by a signed agreement
Make human motion capture easier.
At a glance
- What is it?
- A markerless human motion capture toolbox from the Zhejiang University 3D vision group that fits SMPL family models to RGB video from several input sources. The fitting code is concrete, but the dataset sits behind a signed agreement sent by email and one of the six demo settings is marked coming soon, with one novel view synthesis result captioned the same way.
- Who is it for?
- EasyMocap fits the bill if you already have calibrated cameras or an internet clip and want SMPL family output rather than a bare skeleton, and if you are willing to treat the fitting code as demos you will edit.
- Can I use it commercially?
- Check first. The repository uses a licence we do not classify automatically, so read its LICENSE file before any commercial use.
- Is it still maintained?
- Yes. The repository last received commits 61 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- zju3dv/EasyMocap4,981 ★ · +107
| Repository | 2026-09-23 | 2026-11-15 |
|---|---|---|
| zju3dv/EasyMocap | 4.9k | 5k |
107 Stars in the last 53 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
Four body models, one fitting path
The core of EasyMocap is parametric body fitting, and the first demo listed is multiple views of a single person. That path accepts four models, SMPL, SMPL+H, SMPL-X and MANO, and it targets body, hand and face poses together rather than a bare skeleton. The body model is your first decision. The sample data behind the demos came from ZJU-MoCap, recorded with 23 calibrated and synchronized cameras, and a second set was captured with 8 cameras. Each result is published as a video with a caption naming the camera count, and no accuracy numbers sit next to them, so settings cannot be compared from the repository alone.
Installing means reading a page outside the repository
The Installation section is a single line pointing at an external docs page, https://chingswy.github.io/easymocap-public-doc/install/install.html. No install command sits in the repository, so the dependency story lives off the page you cloned. What you can check locally is requirements.txt, and it is not a light list. Four lines pin an exact version, and one dependency installs straight from a Git URL:
ipdb
joblib
tqdm
opencv-python
yacs
tabulate
termcolor
git+https://github.com/mattloper/chumpy.git
mediapipe==0.10.0
func_timeout
ultralytics
gdown
setuptools==59.5.0
tensorboard==2.8.0
pytorch-lightning==1.5.0mediapipe==0.10.0, setuptools==59.5.0, tensorboard==2.8.0 and pytorch-lightning==1.5.0 are exact pins with no ranges to widen, and the chumpy entry reaches a Git repository rather than a package index. Meanwhile setup.py declares install_requires=[], so an install driven by the package alone pulls nothing and leaves ipdb, opencv-python and the mediapipe pin unresolved. Decide which of the two files is your contract before you start, because they disagree, and expect to resolve the Git dependency yourself.
One console script ships and three sit commented out
setup.py registers exactly one console script, and the other three candidates are commented out in the same block, pointing at utilities you would otherwise find by browsing directories:
entry_points={
'console_scripts': [
'emc=apps.mocap.run:main_entrypoint',
# 'easymocap_calib=easymocap.mytools.entry:calib',
# 'easymocap_tools=easymocap.mytools.entry:main',
# 'extract_keypoints=easymocap.mytools.cmdtools.extract_keypoints:main'
],
},That tells you two things. The calibration and keypoint entry points are named but not installed, so you enable them by editing this block and reinstalling rather than by typing a second command. And the one script that does ship, emc, resolves to apps.mocap.run:main_entrypoint, a path outside the nine packages setup.py lists, which cover easymocap with its config, dataset, smplmodel, pyfitting, mytools, annotator and estimator subpackages, plus myeasymocap, and never apps. So emc resolves from a checkout where apps/ sits beside easymocap/, and the packaged list alone does not carry the command target.
The estimator you pick decides what your footage must contain
Four wrappers under easymocap/estimator form the front end, and each has a stated job:
easymocap/estimator/mediapipe_wrapper.py
easymocap/estimator/SPIN
easymocap/estimator/YOLOv4
easymocap/estimator/HRNetmediapipe_wrapper.py wraps MediaPipe for keypoints, SPIN is an SMPL estimator, YOLOv4 is an object detector, and HRNet is a 2D human pose estimator. That division explains why the demos need different inputs. The internet video path fits SMPL with 2D keypoint estimation plus CNN initialization, so it needs a keypoint estimator and a CNN initializer. The multiple views of multiple people path adds bounding boxes over several subjects, which is the detector's part, and its demo was captured with 8 consumer cameras. A mediapipe keypoints detector landed on 12/25/2021 in the Update list, so the estimator set moved after the tagged release. Choose the wrapper that matches your footage, since the fitting stage downstream consumes whatever the estimator emits.
Two datasets sit behind a signed PDF and two email addresses
The demos run on ZJU-MoCap, two datasets named LightStage and Mirrored-Human, and neither can be fetched from the repository. The route runs through a PDF agreement you sign, then email to Qing Shuai at [email protected] with Xiaowei Zhou at [email protected] copied, and the download link comes back in a reply. If your plan depends on those datasets, a person's inbox sits between you and the first run. The two sets also answer different questions, since LightStage was captured with a LightStage system and Mirrored-Human was collected from the Internet.
Two of the six demo settings are marked as unfinished
Six settings appear under Core features, and one of them is marked coming soon. The entry for multiple internet videos with a specific action carries a Coming soon marker and links a todo document, and the novel view synthesis entry captions one of its results as coming soon as well. The action setting is the gap worth noting, because it is the only one asking several separate clips to agree on a single action. Novel view synthesis is also pointed outward rather than shipped whole, since its link goes to the neuralbody repository, and support for MultiNeuralBody was added on 11/03/2022, the newest entry in the Update list. Read the feature list as a roadmap with open slots, not as six working capabilities.
Calibration and annotation are separate apps with their own pages
Everything outside fitting is split into its own app with its own page. Calibration for intrinsic and extrinsic parameters lives at apps/calibration/Readme.md, the annotator for bounding box, keypoints and mask lives at apps/annotation/Readme.md, and 3D realtime visualization has a page at doc/realtime_visualization.md. The Update list dates the calibration tool and the annotator to 04/11/2021 and realtime visualization to 06/10/2021. The top level shows how much scaffolding sits around the fitting code, and the second requirements file signals that the neural body work carries its own dependency set:
.gitignore
3rdparty/
LICENSE
Readme.md
apps/
config/
data/
doc/
easymocap/
library/
logo.png
myeasymocap/
requirements.txt
requirements_neuralbody.txt
scripts/
setup.pyVersion 0.2.1 in setup.py against a 2021 release tag
Two version signals disagree, and the difference matters before you pin anything. setup.py declares version='0.2.1', and the News section announces EasyMocap v0.2 with support for motion capture from internet videos, yet the only tagged release is v0.1 from 2021-03-29. Commits continue past that, since the last push is 2026-09-15. So there is work in the tree that no release tag covers, and the version in the packaging metadata will not tell you which commit you hold. The license field reads NOASSERTION, a placeholder rather than a license name, so the terms have to be read from the LICENSE file at the top level. Add the upstream sources the credits name, SMPL and SMPL-X layers from MPII, with functions borrowed from SPIN, VIBE and SMPLify-X, and the terms you ship with reach further than this repository.
Editorial conclusion
EasyMocap fits the bill if you already have calibrated cameras or an internet clip and want SMPL family output rather than a bare skeleton, and if you are willing to treat the fitting code as demos you will edit. Before committing, confirm on the external install page that the current pins in requirements.txt still work, read the LICENSE file, since the license field reads NOASSERTION, and start the dataset request early, because the link arrives by email after a signed agreement. If you need a tagged, versioned artifact with a settled command line, the single 2021 release will not give you that.
Frequently asked questions
Is motion capture still used, and does EasyMocap assume it is?
The project is written as current work rather than as an archive: it fits SMPL, SMPL+H, SMPL-X or MANO to RGB video, and the last push on the repository is 2026-09-15. Whether optical marker capture is still in use is a question the project does not answer.
Is there any free motion capture software available inside EasyMocap?
The code is open source and setup.py declares install_requires=[], so the package metadata pulls nothing on its own. The real cost sits in requirements.txt, which pins mediapipe==0.10.0, setuptools==59.5.0, tensorboard==2.8.0 and pytorch-lightning==1.5.0 and installs chumpy from a Git URL, while the license field reads NOASSERTION rather than a license name.
Can you do motion capture with a phone using EasyMocap?
What the demos show is single camera capture from internet video by fitting SMPL with 2D keypoint estimation and CNN initialization, a mirrored video path, and a multi-person multi-view demo captured with 8 consumer cameras. A phone recording would fall closest to the internet video case.
Which AI tool is best for motion capture inside EasyMocap?
EasyMocap bundles choices instead of picking one: easymocap/estimator/mediapipe_wrapper.py for keypoints, easymocap/estimator/SPIN as an SMPL estimator, easymocap/estimator/YOLOv4 as an object detector and easymocap/estimator/HRNet as a 2D human pose estimator. The fitting method is described as similar to SMPLify-X with a 3D keypoints loss, and to TotalCapture without using point clouds.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/zju3dv-easymocap)Star history chart for your README
[](https://hysenlabs.com/projects/zju3dv-easymocap)Was this useful?