MediaPipe Holistic Landmarker
Unified whole-body MediaPipe bundle (holistic_landmarker.task, 13.7 MB, FP16 weights) detecting 553 human body landmarks in real time (33 pose + 478 face/iris + 2Γ21 hand keypoints), plus optional 52 face blendshapes and foreground pose segmentation mask, for the MediaPipe Holistic Landmarker (HolisticLandmarker) task.
holistic_landmarker.task integrates 7 .tflite sub-models across three regions:
- Pose stage:
pose_detector.tflite ([1, 224, 224, 3]) + pose_landmarks_detector.tflite ([1, 256, 256, 3]) -> 33 2D/3D pose landmarks and [1, 256, 256, 1] segmentation mask.
- Face stage:
face_detector.tflite ([1, 128, 128, 3]) + face_landmarks_detector.tflite ([1, 192, 192, 3] lite variant -> 478 face/iris landmarks, distinct from the standalone 256Γ256 face landmarker) + face_blendshapes.tflite ([1, 146, 2] -> 52 blendshape scores).
- Hand stage:
hand_roi_refinement.tflite ([1, 256, 256, 3] -> [1, 1, 1, 4] center + guide points refined from pose wrist/hand keypoints; replaces standalone palm detection) + hand_landmarks_detector.tflite ([1, 224, 224, 3] -> 21 3D landmarks & handedness per hand).
Files
| File |
Precision |
Size |
SHA-256 |
holistic_landmarker.task |
float16 |
13.68 MB (13,683,609 B) |
e2dab61191e2dcd0a15f943d8e3ed1dce13c82dfa597b9dd39f562975a50c3f8 |
Companion Files in Repo Root
Download
pip install -U huggingface_hub
hf download litert-community/MediaPipe-HolisticLandmarker holistic_landmarker.task --local-dir .
Inputs, Outputs, and Model Specifications
Bundle Contents (holistic_landmarker.task)
| Bundled Member |
Uncompressed Size |
SHA-256 (prefix) |
hand_landmarks_detector.tflite |
5.48 MB (5,478,917 B) |
11c272b891e1a99a⦠|
hand_roi_refinement.tflite |
88.4 KB (88,382 B) |
d40b15e15f93f6c9β¦ |
face_detector.tflite |
229.7 KB (229,714 B) |
bbff11cebd1eb27a⦠|
face_landmarks_detector.tflite |
1.15 MB (1,152,950 B) |
bc5ee5de06d8c3a5β¦ |
face_blendshapes.tflite |
955.3 KB (955,312 B) |
4f36dded049db18d⦠|
pose_detector.tflite |
2.96 MB (2,959,046 B) |
9ba9dd3d42efaaba⦠|
pose_landmarks_detector.tflite |
2.82 MB (2,818,358 B) |
1150dc68a713b806β¦ |
Sub-model hand_landmarks_detector.tflite
- Embedded
TFLITE_METADATA: name HandLandmarkDetector Β· packed associated files: handedness.txt
| Direction |
Tensor |
Shape |
Type |
Notes |
| Input |
input_1 |
[1, 224, 224, 3] |
float32 |
image β Input image to be detected. β Normalization: mean=[0.0, 0.0, 0.0], std=[255.0, 255.0, 255.0] |
| Output |
Identity |
[1, 63] |
float32 |
landmarks β The hand landmarks in normalized coordinates. |
| Output |
Identity_1 |
[1, 1] |
float32 |
presense score β The score of hand presence in the image. |
| Output |
Identity_2 |
[1, 1] |
float32 |
handedness β The score of left or right handedness. β Labels: handedness.txt |
| Output |
Identity_3 |
[1, 63] |
float32 |
world landmarks β The hand landmarks in world coordinates. |
Sub-model hand_roi_refinement.tflite
- Embedded
TFLITE_METADATA: name Hand re-crop model.
| Direction |
Tensor |
Shape |
Type |
Notes |
| Input |
input_1 |
[1, 256, 256, 3] |
float32 |
image β Input image to be detected β Normalization: mean=[0.0], std=[255.0] |
| Output |
Identity |
[1, 1, 1, 4] |
float32 |
landmarks β Center point and guide point. Each point contains two float values (x, y). |
Sub-model face_detector.tflite
- Embedded
TFLITE_METADATA: name Short Range Face Detection
| Direction |
Tensor |
Shape |
Type |
Notes |
| Input |
input |
[1, 128, 128, 3] |
float32 |
image β Input image to be detected β Normalization: mean=[127.5], std=[127.5] |
| Output |
regressors |
[1, 896, 16] |
float32 |
raw boxes/keypoints β Undecoded face bboxes location and keypoints |
| Output |
classificators |
[1, 896, 1] |
float32 |
scores β Scores of the detected bboxes. |
Sub-model face_landmarks_detector.tflite
- Embedded
TFLITE_METADATA: name Face mesh detection model v2
| Direction |
Tensor |
Shape |
Type |
Notes |
| Input |
serving_default_input_28:0 |
[1, 192, 192, 3] |
float32 |
image β Input image to be detected β Normalization: mean=[0.0], std=[255.0] |
| Output |
StatefulPartitionedCall:1 |
[1, 1, 1, 1434] |
float32 |
face_landmarks β Undecoded face landmarks. Each landmark contains x, y, and z coordinates |
| Output |
StatefulPartitionedCall:0 |
[1, 1, 1, 1] |
float32 |
presence β Prersence score of the face. |
Sub-model face_blendshapes.tflite
| Direction |
Tensor |
Shape |
Type |
Notes |
| Input |
serving_default_input_points:0 |
[1, 146, 2] |
float32 |
β |
| Output |
StatefulPartitionedCall:0 |
[52] |
float32 |
β |
Sub-model pose_detector.tflite
- Embedded
TFLITE_METADATA: name Full Range Face Detection
| Direction |
Tensor |
Shape |
Type |
Notes |
| Input |
input_1 |
[1, 224, 224, 3] |
float32 |
image β Input image to be detected β Normalization: mean=[127.5], std=[127.5] |
| Output |
Identity |
[1, 2254, 12] |
float32 |
raw boxes/keypoints β Undecoded face bboxes location and keypoints |
| Output |
Identity_1 |
[1, 2254, 1] |
float32 |
scores β Scores of the detected bboxes. |
Sub-model pose_landmarks_detector.tflite
- Embedded
TFLITE_METADATA: name Pose Landmark
| Direction |
Tensor |
Shape |
Type |
Notes |
| Input |
input_1 |
[1, 256, 256, 3] |
float32 |
image β Input image to be landmarked β Normalization: mean=[0.0], std=[255.0] |
| Output |
Identity |
[1, 195] |
float32 |
landmarks β Pose landmarks |
| Output |
Identity_1 |
[1, 1] |
float32 |
presence β Presence of pose. |
| Output |
Identity_2 |
[1, 256, 256, 1] |
float32 |
segmentation_mask β Segmentation mask for pose. |
| Output |
Identity_3 |
[1, 64, 64, 39] |
float32 |
heat_map β Heat map for pose. |
| Output |
Identity_4 |
[1, 117] |
float32 |
world_landmarks β World landmarks for pose. |
Use with MediaPipe Tasks
Python Quickstart
pip install -U mediapipe huggingface_hub
import mediapipe as mp
from huggingface_hub import hf_hub_download
model_path = hf_hub_download(
repo_id="litert-community/MediaPipe-HolisticLandmarker",
filename="holistic_landmarker.task",
)
BaseOptions = mp.tasks.BaseOptions
HolisticLandmarker = mp.tasks.vision.HolisticLandmarker
HolisticLandmarkerOptions = mp.tasks.vision.HolisticLandmarkerOptions
VisionRunningMode = mp.tasks.vision.RunningMode
options = HolisticLandmarkerOptions(
base_options=BaseOptions(model_asset_path=model_path),
running_mode=VisionRunningMode.IMAGE,
min_face_detection_confidence=0.5,
min_pose_detection_confidence=0.5,
output_face_blendshapes=True,
)
with HolisticLandmarker.create_from_options(options) as runner:
mp_image = mp.Image.create_from_file("input.jpg")
result = runner.detect(mp_image)
print(f"pose={len(result.pose_landmarks)} face={len(result.face_landmarks)} blendshapes={len(result.face_blendshapes)} "
f"left_hand={len(result.left_hand_landmarks)} right_hand={len(result.right_hand_landmarks)}")
Cross-Platform Guides & Samples
Intended Use, Limitations, and Fairness
No combined Holistic evaluation PDF is published; summarized from the three bundled upstream component PDF model cards (note that Holistic swaps standalone palm detection for hand_roi_refinement.tflite and uses a compact 192Γ192 face landmark sub-model rather than the 256Γ256 Face Mesh V2 variant evaluated below):
Model Card BlazePose GHUM 3D.pdf (April 16, 2021) β Lite-class pose backbone evaluated on 1,400 back-facing smartphone images (14 UN subregions: 87.0% avg PDJ; Fitzpatrick 1β6: 86.2%; gender: 86.0%β89.1%).
Model Card Hand Tracking (Lite_Full) with Fairness Oct 2021.pdf (Oct 2021) β Full hand landmarker evaluated across 14 UN subregions (10.09% Β± 1.73% MNAE) and 6 Fitzpatrick groups (5.08% Β± 0.72% MNAE).
Model Card MediaPipe Face Mesh V2.pdf (September 15, 2022) β 478 face landmarks evaluated across 17 UN subregions (2.62% tracking IOD MAE).
- Out-of-scope uses:
- Any form of surveillance or identity recognition is explicitly out of scope and not enabled by this technology.
- Predicted landmarks do not provide facial recognition or identification and do not store any unique face representation.
- Human life-critical decisions, multi-person scenes, people > 4 meters from the camera, or metric-accurate depth.
Provenance
Every model binary in this repository is byte-identical (unmodified weights, metadata, and .task bundle archive bytes) to the official artifact served by Google from storage.googleapis.com/mediapipe-models.
| Repo File |
Upstream URL |
GCS Generation |
Upstream Last-Modified |
Pinned Bucket Version |
holistic_landmarker.task |
holistic_landmarker.task |
1703178474695092 |
Thu, 21 Dec 2023 17:07:54 GMT |
/1/ |
License and Citation
- License:
apache-2.0 β Stated as Apache License, Version 2.0 in the bundled upstream Google model-card PDFs.
Citation
@article{bazarevsky2020blazepose,
title = {BlazePose: On-device Real-time Body Pose tracking},
author = {Bazarevsky, Valentin and Grishchenko, Ivan and Raveendran, Karthik and Zhu, Tyler and Zhang, Fan and Grundmann, Matthias},
journal = {arXiv preprint arXiv:2006.10204},
year = {2020},
url = {https://arxiv.org/pdf/2006.10204}
}
@inproceedings{xu2020ghum,
title = {GHUM & GHUML: Generative 3D Human Shape and Articulated Pose Models},
author = {Xu, Hongyi and Bazavan, Eduard Gabriel and Zanfir, Andrei and Freeman, William T. and Sukthankar, Rahul and Sminchisescu, Cristian},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
pages = {6184--6193},
year = {2020}
}
@article{zhang2020mediapipehands,
title = {MediaPipe Hands: On-device Real-time Hand Tracking},
author = {Zhang, Fan and Bazarevsky, Valentin and Vakunov, Andrey and Tkachenka, Andrei and Sung, George and Chang, Chuo-Ling and Grundmann, Matthias},
journal = {arXiv preprint arXiv:2006.10214},
year = {2020},
url = {https://arxiv.org/abs/2006.10214}
}