MediaPipe Hand Landmarker (Full)

Two-stage MediaPipe task bundle (hand_landmarker.task, 7.8 MB, FP16 weights) combining a palm detector and a 21-keypoint 3D hand landmark regression model for the MediaPipe Hand Landmarker (HandLandmarker) task across Web, Android, iOS, Python, and Flutter.

Documentation: MediaPipe Hand Landmarker Docs

Samples: Android · iOS · Python Colab · Web · Flutter

Try live: Web Demo

hand_landmarker.task packages both neural networks required by HandLandmarker into a single bundle:

  1. Palm detection model (hand_detector.tflite): Single-shot detector that locates hands across the full input frame.
  2. Hand landmarks model (hand_landmarks_detector.tflite): Regresses 21 3D hand-knuckle coordinates in normalized image space, 21 3D metric-scale world coordinates (estimated via the articulated GHUM hand model), a hand-presence score, and subject handedness (Left / Right) on the cropped hand region.

Trained on approximately 30K real-world images alongside rendered synthetic hand poses over varied backgrounds. In VIDEO and LIVE_STREAM running modes, HandLandmarker reuses the previous frame's landmark bounding box to localize hands on subsequent frames and only re-triggers palm detection when hand presence or tracking confidence falls below threshold.

Handedness note: Left and Right handedness correspond to the subject's hands, not the viewer (mirrored selfie frames should be flipped or coordinates inverted).

21 Hand Landmark Indices

ID Landmark ID Landmark
0 WRIST 11 MIDDLE_FINGER_DIP
1 THUMB_CMC 12 MIDDLE_FINGER_TIP
2 THUMB_MCP 13 RING_FINGER_MCP
3 THUMB_IP 14 RING_FINGER_PIP
4 THUMB_TIP 15 RING_FINGER_DIP
5 INDEX_FINGER_MCP 16 RING_FINGER_TIP
6 INDEX_FINGER_PIP 17 PINKY_MCP
7 INDEX_FINGER_DIP 18 PINKY_PIP
8 INDEX_FINGER_TIP 19 PINKY_DIP
9 MIDDLE_FINGER_MCP 20 PINKY_TIP
10 MIDDLE_FINGER_PIP

Files

File Precision Size SHA-256
hand_landmarker.task float16 7.82 MB (7,819,105 B) fbc2a30080c3c557093b5ddfc334698132eb341044ccee322ccf8bcf3607cde1

Companion Files in Repo Root

File Role Size SHA-256
Model Card Hand Tracking (Lite_Full) with Fairness Oct 2021.pdf Upstream Google model card (PDF) 358.0 KB (358,044 B) 43127ff8a92e22717f0d8c30dab3efef4641a29cbf50b15ae3de001be94ab76b

Download

pip install -U huggingface_hub
hf download litert-community/MediaPipe-HandLandmarker-Full hand_landmarker.task --local-dir .

Inputs, Outputs, and Model Specifications

Bundle Contents (hand_landmarker.task)

Bundled Member Uncompressed Size SHA-256 (prefix)
hand_detector.tflite 2.34 MB (2,339,878 B) 945f713bc23570bd…
hand_landmarks_detector.tflite 5.48 MB (5,478,949 B) 6acda74af3fbf40e…

Sub-model hand_detector.tflite

  • Embedded TFLITE_METADATA: name Multi Hand Detection · version 1 · author MediaPipe · packed associated files: labels.txt
Direction Tensor Shape Type Notes
Input input_1 [1, 192, 192, 3] float32 image — Input image to be detected — Normalization: mean=[0.0], std=[255.0]
Output Identity [1, 2016, 18] float32 palm detection — regressed detected palm locations
Output Identity_1 [1, 2016, 1] float32 scores — palm detection scores — Labels: labels.txt

Sub-model hand_landmarks_detector.tflite

  • Embedded TFLITE_METADATA: name HandLandmarkDetector · version 1 · author MediaPipe · packed associated files: handedness.txt
Direction Tensor Shape Type Notes
Input input_1 [1, 224, 224, 3] float32 image — Input image to be detected. — Normalization: mean=[0.0, 0.0, 0.0], std=[255.0, 255.0, 255.0]
Output Identity [1, 63] float32 landmarks — The hand landmarks in normalized coordinates.
Output Identity_1 [1, 1] float32 presense score — The score of hand presence in the image.
Output Identity_2 [1, 1] float32 handedness — The score of left or right handedness. — Labels: handedness.txt
Output Identity_3 [1, 63] float32 world landmarks — The hand landmarks in world coordinates.

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-HandLandmarker-Full",
    filename="hand_landmarker.task",
)

BaseOptions = mp.tasks.BaseOptions
HandLandmarker = mp.tasks.vision.HandLandmarker
HandLandmarkerOptions = mp.tasks.vision.HandLandmarkerOptions
VisionRunningMode = mp.tasks.vision.RunningMode

options = HandLandmarkerOptions(
    base_options=BaseOptions(model_asset_path=model_path),
    running_mode=VisionRunningMode.IMAGE,
    num_hands=2,
)

with HandLandmarker.create_from_options(options) as runner:
    mp_image = mp.Image.create_from_file("input.jpg")
    result = runner.detect(mp_image)
    for handedness, landmarks in zip(result.handedness, result.hand_landmarks):
        side = handedness[0].category_name
        wrist = landmarks[0]
        print(f"{side} hand: wrist=({wrist.x:.3f}, {wrist.y:.3f}, {wrist.z:.3f})")

Cross-Platform Guides & Samples

Latency Benchmarks

End-to-end MediaPipe Task pipeline average latency reported by the Hand Landmarker documentation:

Model Variant Device CPU Latency GPU Latency
HandLandmarker (full) Pixel 6 17.12 ms 12.27 ms

Intended Use, Limitations, and Fairness

Summarized from the bundled Model Card Hand Tracking (Lite_Full) with Fairness Oct 2021.pdf (Oct 2021):

  • Intended applications: Predicting 2D screen and 3D world landmarks on prominently displayed hands in smartphone or webcam images and video streams for mobile AR, gesture recognition, and hand-control interfaces.
  • Out-of-scope uses:
    • Any form of surveillance or identity recognition is explicitly out of scope and not enabled by this technology.
    • Human life-critical decisions.
    • Counting hands in crowds, or predicting landmarks on hands wearing gloves, holding occluding objects, or covered by heavy decorations (jewelry, tattoos, henna).
  • Known limitations: Trained on finite real and synthetic datasets; blur, heavy self-occlusion, low light, and low-end camera noise increase landmark error (with error smallest near the rigid palm/MCP joints and largest at the fingertips).
  • Fairness evaluation (Full model summary): Measured via palm-size-normalized Mean Absolute Error (MNAE) on 2D screen landmarks:
    • 14 UN geographic subregions (700 images, 50/region): average MNAE 10.09% ± 1.73% (range 6.10%–13.00%, vs. 6.0% human inter-annotator baseline).
    • 6 Fitzpatrick skin-tone types (420 images): average MNAE 5.08% ± 0.72% (range 4.53%–6.21%, vs. 3.8% human inter-annotator baseline).
    • Perceived gender: average MNAE 4.80% (female) and 5.38% (male). Refer to the bundled PDF model card for full regional and demographic breakdowns.

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
hand_landmarker.task hand_landmarker.task 1682480005356399 Wed, 26 Apr 2023 03:33:25 GMT /1/

License and Citation

  • License: apache-2.0 — Stated as Apache License, Version 2.0 in the bundled upstream Google model-card PDF.

Citation

@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}
}
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Collection including litert-community/MediaPipe-HandLandmarker-Full

Paper for litert-community/MediaPipe-HandLandmarker-Full