Project Overview
MediaPipe is Google's open-source framework for running machine learning pipelines on live and streaming media. You point it at a camera feed, a video file, or an audio stream, and it gives you real-time hand tracking, pose estimation, face landmarks, segmentation, and object detection — on Android, iOS, the web, and desktop, with the same models running everywhere.
It's really two products in one repo. The first is a set of prebuilt, pretrained tasks you call from Python, JavaScript, Kotlin, Swift, or C++ without training anything. The second is a lower-level C++ graph framework for wiring custom perception pipelines out of reusable components. Most teams only ever touch the first layer, and that's the right call — the prebuilt tasks are the main product.
The license is Apache 2.0, so commercial use is straightforward. One structural shift matters before you write any code: the original Solutions API (the mp.solutions namespace you'll see in older tutorials) is in maintenance mode, and active development happens in the MediaPipe Tasks API. Start with Tasks and treat pre-2023 tutorials with suspicion. The official documentation lives at ai.google.dev/edge/mediapipe.
Key Challenges Addressed
Real-time perception is hard to ship, and MediaPipe targets the specific problems that make it hard:
- On-device inference speed. Models run through TensorFlow Lite with GPU delegates — OpenGL ES on Android, Metal on iOS — so you get usable frame rates on midrange phones without a server round trip or per-frame API costs.
- Frame synchronization and backpressure. The graph runtime stamps every packet with a timestamp and drops frames when a downstream stage can't keep up, instead of letting queues grow without bound. Latency stays flat under load rather than drifting seconds behind the camera.
- Cross-platform duplication. A pipeline is defined once as a graph of calculators and runs on every supported platform. You don't rewrite the hand tracker for iOS after shipping it on Android.
- Cold start. Pretrained models for hands, pose, faces, and segmentation mean you skip data collection and training entirely for the common cases.
Getting Started
Python is the fastest way in:
pip install mediapipe
Download a model bundle for your task, then run it. Hand landmarks look like this:
import mediapipe as mp
from mediapipe.tasks import python
from mediapipe.tasks.python import vision
base_options = python.BaseOptions(model_asset_path='hand_landmarker.task')
options = vision.HandLandmarkerOptions(base_options=base_options, num_hands=2)
detector = vision.HandLandmarker.create_from_options(options)
image = mp.Image.create_from_file('hand.jpg')
result = detector.detect(image)
print(result.hand_landmarks)
Sharp edges you'll hit in the first hour:
- If
pip install mediapipefails with "No matching distribution found", your Python version is newer than the published wheels. Create a virtualenv on a supported Python version first instead of fighting pip. create_from_optionsraises at startup ifmodel_asset_pathdoesn't point to a real.taskbundle — the model is a separate download, not part of the pip package. Grab the bundle from the official model page for your task before running anything.- Mixing APIs breaks immediately:
mp.solutions.handssnippets from old tutorials won't work with Tasks objects. Pick the Tasks API and ignore code that importsmp.solutions. - In live-stream mode, sending frames with non-increasing timestamps makes the graph error out and stop. Derive timestamps from a monotonic counter, never from wall-clock time that can repeat.
- Building the C++ framework from source requires the exact Bazel version the repo pins. Use Bazelisk, which reads
.bazelversionautomatically, instead of a system Bazel that will fail mid-build.
Features and Use Cases
The Tasks API groups features by domain:
- Vision: hand landmarks (21 keypoints per hand), pose landmarks (33 keypoints), face landmarks (478 points plus blendshape scores), gesture recognition, image and selfie segmentation, object detection, image classification, and image embeddings.
- Audio: audio classification.
- Text: text classification and language detection.
- GenAI: an LLM inference API for running small language models on-device.
What that translates to in practice:
- AR effects and try-on. Face landmarks plus blendshape scores drive filters, makeup try-on, and avatar animation from a selfie camera.
- Fitness and movement apps. Pose landmarks give you joint positions per frame; you build rep counting and form checks on top.
- Gesture control. The gesture recognizer ships with common gestures, and Model Maker lets you retrain it on your own gesture set without touching model internals.
- Background replacement. Selfie segmentation produces a per-pixel mask fast enough for live video calls.
Model Maker is the customization story: transfer learning for a handful of task types, gesture recognition and image classification among them. If your need falls outside what it supports, you're into full custom-model territory — budget for that before you commit.
Ecosystem and Dependencies
MediaPipe runs inference through TensorFlow Lite, which lives in the TensorFlow repo. Desktop C++ examples use OpenCV for camera capture and window display. On the web, tasks run through WebAssembly with WebGL acceleration via the @mediapipe/tasks-vision npm package — no plugin required. Source builds use Bazel.
The practical gain: prototype in Python, then ship the same .task model bundle to Android (Kotlin), iOS (Swift), and the browser (JavaScript) without conversion work.
Architectural Overview
Every MediaPipe pipeline is a graph. Nodes are calculators — C++ classes that consume and produce typed packets. Edges are streams of timestamped packets. Graph topology lives in .pbtxt text protos, separate from calculator code.
Three design choices matter for you:
- Timestamped packets let the framework synchronize multiple streams (video frames plus detection results) and drop stale work deliberately. That's the mechanism behind stable live-stream latency.
- Calculators are reusable and individually testable. The prebuilt tasks are compositions of the same calculators you'd use in a custom graph, so customization extends the system rather than forking it.
- Graph config as data means pipeline changes don't require recompiling app code — but errors surface as C++ status messages at graph initialization, which is painful to debug from Python or JavaScript.
The cost of the design is the cliff between consuming prebuilt tasks (easy) and authoring custom calculators (C++, Bazel, protobuf). Know which side of that cliff your project sits on before committing.
Pros and Cons
Pros:
- Real-time performance on phones is the headline: GPU-accelerated inference designed for 30fps live streams.
- One model bundle and consistent APIs across Android, iOS, web, Python, and C++.
- Pretrained models under Apache 2.0 — no training infrastructure, no licensing surprises.
- The Tasks API gets you working results in about twenty lines of Python.
Cons:
- Customization beyond Model Maker's supported task types is a steep drop into C++ graphs and Bazel builds.
- Search results and Stack Overflow answers skew heavily toward the deprecated Solutions API, which wastes debugging time.
- Desktop Python inference is CPU-oriented; GPU acceleration is strongest on mobile and web.
- You can't retrain the core models — the model zoo updates on Google's schedule, not yours.
Comparison and Alternatives
GitHub stars
Weekly star gains
Full metrics details
GitHub GraphQL current stargazer cohort, reconstructed from starredAt events and totalCount retrieved in one stable pagination walk. GitHub does not expose historical unstar events, so this is not an exact ledger of past star counts.
Mediapipe
GitHub GraphQL current stargazer cohort, reconstructed from starredAt events and totalCount retrieved in one stable pagination walk. GitHub does not expose historical unstar events, so this is not an exact ledger of past star counts.
180 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
OpenPose
GitHub GraphQL current stargazer cohort, reconstructed from starredAt events and totalCount retrieved in one stable pagination walk. GitHub does not expose historical unstar events, so this is not an exact ledger of past star counts.
180 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
OpenCV
GitHub GraphQL current stargazer cohort, reconstructed from starredAt events and totalCount retrieved in one stable pagination walk. GitHub does not expose historical unstar events, so this is not an exact ledger of past star counts.
80 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
tfjs-models
GitHub GraphQL current stargazer cohort, reconstructed from starredAt events and totalCount retrieved in one stable pagination walk. GitHub does not expose historical unstar events, so this is not an exact ledger of past star counts.
180 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
Issues opened
Weekly total
Full metrics details
GitHub API observation. Historical values render only when the source provides a real observation for that day.
Mediapipe
GitHub API observation. Historical values render only when the source provides a real observation for that day.
144 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
OpenPose
GitHub API observation. Historical values render only when the source provides a real observation for that day.
144 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
OpenCV
GitHub API observation. Historical values render only when the source provides a real observation for that day.
36 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
tfjs-models
GitHub API observation. Historical values render only when the source provides a real observation for that day.
143 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
Issues closed
Weekly total
Full metrics details
GitHub Search issue totals queried for exact UTC days; rolling values are sums of proven daily counts.
Mediapipe
GitHub Search issue totals queried for exact UTC days; rolling values are sums of proven daily counts.
144 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
OpenPose
GitHub Search issue totals queried for exact UTC days; rolling values are sums of proven daily counts.
144 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
OpenCV
GitHub Search issue totals queried for exact UTC days; rolling values are sums of proven daily counts.
36 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
tfjs-models
GitHub Search issue totals queried for exact UTC days; rolling values are sums of proven daily counts.
143 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
Pull requests opened
Weekly total
Full metrics details
GitHub API observation. Historical values render only when the source provides a real observation for that day.
Mediapipe
GitHub API observation. Historical values render only when the source provides a real observation for that day.
144 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
OpenPose
GitHub API observation. Historical values render only when the source provides a real observation for that day.
144 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
OpenCV
GitHub API observation. Historical values render only when the source provides a real observation for that day.
36 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
tfjs-models
GitHub API observation. Historical values render only when the source provides a real observation for that day.
143 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
Pull requests closed
Weekly total
Full metrics details
GitHub Search pull-request totals queried for exact UTC days; rolling values are sums of proven daily counts.
Mediapipe
GitHub Search pull-request totals queried for exact UTC days; rolling values are sums of proven daily counts.
144 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
OpenPose
GitHub Search pull-request totals queried for exact UTC days; rolling values are sums of proven daily counts.
144 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
OpenCV
GitHub Search pull-request totals queried for exact UTC days; rolling values are sums of proven daily counts.
36 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
tfjs-models
GitHub Search pull-request totals queried for exact UTC days; rolling values are sums of proven daily counts.
143 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
Pull requests merged
Weekly total
Full metrics details
GitHub API observation. Historical values render only when the source provides a real observation for that day.
Mediapipe
GitHub API observation. Historical values render only when the source provides a real observation for that day.
144 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
OpenPose
GitHub API observation. Historical values render only when the source provides a real observation for that day.
144 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
OpenCV
GitHub API observation. Historical values render only when the source provides a real observation for that day.
36 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
tfjs-models
GitHub API observation. Historical values render only when the source provides a real observation for that day.
143 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
Open/closed pull request ratio
Weekly average
Full metrics details
GitHub Search pull-request totals queried for exact UTC days; rolling values are sums of proven daily counts.
Mediapipe
GitHub Search pull-request totals queried for exact UTC days; rolling values are sums of proven daily counts.
180 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
OpenPose
GitHub Search pull-request totals queried for exact UTC days; rolling values are sums of proven daily counts.
180 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
OpenCV
GitHub Search pull-request totals queried for exact UTC days; rolling values are sums of proven daily counts.
80 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
tfjs-models
GitHub Search pull-request totals queried for exact UTC days; rolling values are sums of proven daily counts.
180 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
Open/closed issues ratio
Weekly average
Full metrics details
GitHub Search issue totals queried for exact UTC days; rolling values are sums of proven daily counts.
Mediapipe
GitHub Search issue totals queried for exact UTC days; rolling values are sums of proven daily counts.
180 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
OpenPose
GitHub Search issue totals queried for exact UTC days; rolling values are sums of proven daily counts.
180 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
OpenCV
GitHub Search issue totals queried for exact UTC days; rolling values are sums of proven daily counts.
80 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
tfjs-models
GitHub Search issue totals queried for exact UTC days; rolling values are sums of proven daily counts.
180 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
Releases
Weekly total
Full metrics details
GitHub release published_at events bucketed by UTC day; drafts are excluded.
Mediapipe
GitHub release published_at events bucketed by UTC day; drafts are excluded.
180 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
OpenPose
GitHub release published_at events bucketed by UTC day; drafts are excluded.
180 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
OpenCV
GitHub release published_at events bucketed by UTC day; drafts are excluded.
80 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
tfjs-models
GitHub release published_at events bucketed by UTC day; drafts are excluded.
180 observed daily rows. Missing days are not fabricated.
| Date | Value |
|---|
Hotness score
Weekly average
Full metrics details
Derived only from GitHub GraphQL starredAt events after daily star totals are reconciled.
Mediapipe
Derived only from GitHub GraphQL starredAt events after daily star totals are reconciled.
180 observed daily rows. Missing days are not fabricated.
Per-day formula: 0.40 × Hot today + 0.40 × Hot week + 0.20 × Breakout.
- Same-day multiplier = stars 1d ÷ max(1, stars 7d ÷ 7)
- Weekly multiplier = stars 7d ÷ max(1, stars 30d ÷ 30 × 7)
- Fortnight multiplier = stars 14d ÷ max(1, stars 90d ÷ 90 × 14)
- Hot today = clamp(70 × log-scale(stars 1d, 1000) + 30 × breakout-scale(same-day, 4))
- Hot week = clamp(75 × log-scale(stars 7d, 5000) + 25 × breakout-scale(weekly, 3))
- Breakout = clamp(35 × breakout-scale(same-day, 4) + 40 × breakout-scale(weekly, 3) + 25 × breakout-scale(fortnight, 2.5))
| Date | Stars | Stars 1d | Stars 7d | Stars 14d | Stars 30d | Stars 90d | Same-day multiplier | Weekly multiplier | Fortnight multiplier | Hot today | Hot week | Breakout | Stored score |
|---|
OpenPose
Derived only from GitHub GraphQL starredAt events after daily star totals are reconciled.
180 observed daily rows. Missing days are not fabricated.
Per-day formula: 0.40 × Hot today + 0.40 × Hot week + 0.20 × Breakout.
- Same-day multiplier = stars 1d ÷ max(1, stars 7d ÷ 7)
- Weekly multiplier = stars 7d ÷ max(1, stars 30d ÷ 30 × 7)
- Fortnight multiplier = stars 14d ÷ max(1, stars 90d ÷ 90 × 14)
- Hot today = clamp(70 × log-scale(stars 1d, 1000) + 30 × breakout-scale(same-day, 4))
- Hot week = clamp(75 × log-scale(stars 7d, 5000) + 25 × breakout-scale(weekly, 3))
- Breakout = clamp(35 × breakout-scale(same-day, 4) + 40 × breakout-scale(weekly, 3) + 25 × breakout-scale(fortnight, 2.5))
| Date | Stars | Stars 1d | Stars 7d | Stars 14d | Stars 30d | Stars 90d | Same-day multiplier | Weekly multiplier | Fortnight multiplier | Hot today | Hot week | Breakout | Stored score |
|---|
OpenCV
Derived only from GitHub GraphQL starredAt events after daily star totals are reconciled.
80 observed daily rows. Missing days are not fabricated.
Per-day formula: 0.40 × Hot today + 0.40 × Hot week + 0.20 × Breakout.
- Same-day multiplier = stars 1d ÷ max(1, stars 7d ÷ 7)
- Weekly multiplier = stars 7d ÷ max(1, stars 30d ÷ 30 × 7)
- Fortnight multiplier = stars 14d ÷ max(1, stars 90d ÷ 90 × 14)
- Hot today = clamp(70 × log-scale(stars 1d, 1000) + 30 × breakout-scale(same-day, 4))
- Hot week = clamp(75 × log-scale(stars 7d, 5000) + 25 × breakout-scale(weekly, 3))
- Breakout = clamp(35 × breakout-scale(same-day, 4) + 40 × breakout-scale(weekly, 3) + 25 × breakout-scale(fortnight, 2.5))
| Date | Stars | Stars 1d | Stars 7d | Stars 14d | Stars 30d | Stars 90d | Same-day multiplier | Weekly multiplier | Fortnight multiplier | Hot today | Hot week | Breakout | Stored score |
|---|
tfjs-models
Derived only from GitHub GraphQL starredAt events after daily star totals are reconciled.
180 observed daily rows. Missing days are not fabricated.
Per-day formula: 0.40 × Hot today + 0.40 × Hot week + 0.20 × Breakout.
- Same-day multiplier = stars 1d ÷ max(1, stars 7d ÷ 7)
- Weekly multiplier = stars 7d ÷ max(1, stars 30d ÷ 30 × 7)
- Fortnight multiplier = stars 14d ÷ max(1, stars 90d ÷ 90 × 14)
- Hot today = clamp(70 × log-scale(stars 1d, 1000) + 30 × breakout-scale(same-day, 4))
- Hot week = clamp(75 × log-scale(stars 7d, 5000) + 25 × breakout-scale(weekly, 3))
- Breakout = clamp(35 × breakout-scale(same-day, 4) + 40 × breakout-scale(weekly, 3) + 25 × breakout-scale(fortnight, 2.5))
| Date | Stars | Stars 1d | Stars 7d | Stars 14d | Stars 30d | Stars 90d | Same-day multiplier | Weekly multiplier | Fortnight multiplier | Hot today | Hot week | Breakout | Stored score |
|---|
Reliability score
Weekly average
Full metrics details
Derived from the displayed continuity, closure, shipping, liveness, and support-burden components.
Mediapipe
Derived from the displayed continuity, closure, shipping, liveness, and support-burden components.
180 observed daily rows. Missing days are not fabricated.
Per-day formula: 0.30 × Continuity + 0.30 × Closure + 0.20 × Shipping + 0.10 × Liveness + 0.10 × Support burden.
- Continuity = 35% active ratio 90d + 20% active ratio 30d + 15% PR efficiency 90d + 10% PR efficiency 30d + 10% liveness + 10% observed-history coverage
- Closure = 55% PR merge efficiency 90d + 45% issue close efficiency 90d
- Shipping = 100 × (20% × release ratio 30d + 30% × release ratio 90d + 50% × release ratio 180d); ratios are releases ÷ 2, 6, and 12, capped at 1
- Support burden = 100 − 2 × open issues per 1,000 stars
- Liveness = 100 × exp(−pushed days ago ÷ 120)
Stars, issues, pull requests, and releases are daily observations. Pushed-days and license are repository snapshot-derived inputs and are not independent historical GitHub events.
| Date | Issues opened | Issues closed | PRs opened | PRs merged | Releases | Stars | Open issues | Pushed days ago | Continuity | Closure | Shipping | Liveness | Support burden | License | Observed days | Missing days | Needs healing | Stored score |
|---|
OpenPose
Derived from the displayed continuity, closure, shipping, liveness, and support-burden components.
180 observed daily rows. Missing days are not fabricated.
Per-day formula: 0.30 × Continuity + 0.30 × Closure + 0.20 × Shipping + 0.10 × Liveness + 0.10 × Support burden.
- Continuity = 35% active ratio 90d + 20% active ratio 30d + 15% PR efficiency 90d + 10% PR efficiency 30d + 10% liveness + 10% observed-history coverage
- Closure = 55% PR merge efficiency 90d + 45% issue close efficiency 90d
- Shipping = 100 × (20% × release ratio 30d + 30% × release ratio 90d + 50% × release ratio 180d); ratios are releases ÷ 2, 6, and 12, capped at 1
- Support burden = 100 − 2 × open issues per 1,000 stars
- Liveness = 100 × exp(−pushed days ago ÷ 120)
Stars, issues, pull requests, and releases are daily observations. Pushed-days and license are repository snapshot-derived inputs and are not independent historical GitHub events.
| Date | Issues opened | Issues closed | PRs opened | PRs merged | Releases | Stars | Open issues | Pushed days ago | Continuity | Closure | Shipping | Liveness | Support burden | License | Observed days | Missing days | Needs healing | Stored score |
|---|
OpenCV
Derived from the displayed continuity, closure, shipping, liveness, and support-burden components.
80 observed daily rows. Missing days are not fabricated.
Per-day formula: 0.30 × Continuity + 0.30 × Closure + 0.20 × Shipping + 0.10 × Liveness + 0.10 × Support burden.
- Continuity = 35% active ratio 90d + 20% active ratio 30d + 15% PR efficiency 90d + 10% PR efficiency 30d + 10% liveness + 10% observed-history coverage
- Closure = 55% PR merge efficiency 90d + 45% issue close efficiency 90d
- Shipping = 100 × (20% × release ratio 30d + 30% × release ratio 90d + 50% × release ratio 180d); ratios are releases ÷ 2, 6, and 12, capped at 1
- Support burden = 100 − 2 × open issues per 1,000 stars
- Liveness = 100 × exp(−pushed days ago ÷ 120)
Stars, issues, pull requests, and releases are daily observations. Pushed-days and license are repository snapshot-derived inputs and are not independent historical GitHub events.
| Date | Issues opened | Issues closed | PRs opened | PRs merged | Releases | Stars | Open issues | Pushed days ago | Continuity | Closure | Shipping | Liveness | Support burden | License | Observed days | Missing days | Needs healing | Stored score |
|---|
tfjs-models
Derived from the displayed continuity, closure, shipping, liveness, and support-burden components.
180 observed daily rows. Missing days are not fabricated.
Per-day formula: 0.30 × Continuity + 0.30 × Closure + 0.20 × Shipping + 0.10 × Liveness + 0.10 × Support burden.
- Continuity = 35% active ratio 90d + 20% active ratio 30d + 15% PR efficiency 90d + 10% PR efficiency 30d + 10% liveness + 10% observed-history coverage
- Closure = 55% PR merge efficiency 90d + 45% issue close efficiency 90d
- Shipping = 100 × (20% × release ratio 30d + 30% × release ratio 90d + 50% × release ratio 180d); ratios are releases ÷ 2, 6, and 12, capped at 1
- Support burden = 100 − 2 × open issues per 1,000 stars
- Liveness = 100 × exp(−pushed days ago ÷ 120)
Stars, issues, pull requests, and releases are daily observations. Pushed-days and license are repository snapshot-derived inputs and are not independent historical GitHub events.
| Date | Issues opened | Issues closed | PRs opened | PRs merged | Releases | Stars | Open issues | Pushed days ago | Continuity | Closure | Shipping | Liveness | Support burden | License | Observed days | Missing days | Needs healing | Stored score |
|---|
If MediaPipe doesn't fit, here's where to look:
- OpenPose is the classic multi-person pose estimator. It's accurate, but it needs a serious GPU, and its license restricts commercial use without a paid agreement — read it before building a product on it.
- OpenCV is the general-purpose computer vision library. It won't hand you a pretrained hand tracker out of the box, but its DNN module runs public models and you control every pipeline stage.
- tfjs-models packages browser-ready models for pose, face, and hand detection — overlapping scope with MediaPipe, but web-only.
- InsightFace goes deeper on faces: recognition and identity, which MediaPipe's face tasks don't attempt.
- PaddleDetection is a training-first detection and keypoint toolkit. Choose it when you must train custom models rather than consume pretrained ones.
| Project | Repo | Hotness (~6mo) | PRs merged (~6mo) | Star growth (~6mo) |
|---|---|---|---|---|
| MediaPipe | https://github.com/google/mediapipe | 12.40 | 0 | 14954 |
| OpenPose | https://github.com/CMU-Perceptual-Computing-Lab/openpose | 6.33 | 0 | 1460 |
| OpenCV | https://github.com/opencv/opencv | 2.42 | 0 | 15221 |
| tfjs-models | https://github.com/tensorflow/tfjs-models | metrics pending | 0 | 9 |
| InsightFace | https://github.com/deepinsight/insightface | metrics pending | 0 | 186 |
| PaddleDetection | https://github.com/PaddlePaddle/PaddleDetection | metrics pending | 0 | metrics pending |
The best overall alternative is OpenCV: it's actively maintained, its documentation is mature, and it runs on every platform MediaPipe targets while leaving you in full control of the models you deploy.
Conclusion
If you're building a mobile or web app that needs real-time hand, pose, or face tracking, use MediaPipe — the pretrained Tasks API gets you a working prototype in an afternoon, and the same model bundles ship to production. Start with the official docs at ai.google.dev/edge/mediapipe and the Python quickstarts.
Do not build new code on the legacy mp.solutions API — it's in maintenance mode, its models no longer receive updates, and code written against it won't migrate cleanly to the Tasks API. Do this first: run pip install mediapipe, download the hand_landmarker.task bundle from the official model page, and run the hand landmarker quickstart end to end.