ML Kit Pose Detection: Base vs Accurate SDK, Landmarks and Limits

Choosing between ML Kit, MediaPipe and the rest happens in pose estimation models compared. ML Kit is governed by service terms rather than an open-source licence, read in pose estimation model licenses. This page covers what Google's ML Kit guides and API reference say about the pose API, as read on October 17, 2026.

Base vs accurate SDK#

Google's overview says the base SDK "runs in real time on modern phones like the Pixel 4 and iPhone X" while "the precision of the landmark coordinates may vary", and that the accurate SDK "returns results at a slower framerate, but produces more accurate coordinate values". On Android the code and assets are statically linked into the app at build time, and the guide requires API level 23 or above; on iOS the detector's assets are statically linked. The guides printed these dependency versions on the day we read them:

com.google.mlkit:pose-detection:18.0.0-beta5
com.google.mlkit:pose-detection-accurate:18.0.0-beta5
pod 'GoogleMLKit/PoseDetection', '8.0.0'
pod 'GoogleMLKit/PoseDetectionAccurate', '8.0.0'

The iOS guide asks for Xcode 13.2.1 or higher.

What it returns#

When a person is detected, the Android guide says the API returns a Pose with 33 PoseLandmark objects, and an empty list when nobody is. Each landmark has a type, a 2D position, a 3D position and an in-frame likelihood from 0.0 to 1.0. When part of the body is outside the image, Google says the missing landmarks get coordinates outside the frame and low in-frame values.

The 33 types are listed with their integer values in the ML Kit landmark map. They use the same 33 values as MediaPipe's PoseLandmark enum, and every name matches except the mouth corners: ML Kit's LEFT_MOUTH and RIGHT_MOUTH against MediaPipe's MOUTH_LEFT and MOUTH_RIGHT. One quirk of Google's Android reference: it describes RIGHT_EYE, RIGHT_EYE_INNER and RIGHT_EYE_OUTER with the word "left".

Google's overview adds that "Pose Detection does not attach any meaning to the landmarks in a pose or the pose itself". Classifying a pose or counting a rep is your code's job, which how rep counting works covers.

The Z value#

Google calls Z "an experimental value". The overview says it is measured in image pixels like X and Y but "is not a true 3D value", that its origin is approximately the center point between the hips, that negative values are toward the camera, and that it has no upper or lower bound. The Android reference adds that Z is less accurate than X and Y.

Two of Google's pages differ on one point. The overview says the Z coordinate "is calculated for every landmark"; the iOS Swift reference says "z-values for facial landmarks are not calculated, so please disregard them". Treat facial Z as unusable either way. For depth you can lean on, see 2D vs 3D pose estimation.

Detection modes and inputs#

STREAM_MODE, the default, detects the most prominent person, then tracks them and re-runs person detection only when they are obscured or lost; Google recommends it for video. SINGLE_IMAGE_MODE runs person detection on every image, with higher latency and no tracking, for static images. On Android, setPreferredHardwareConfigs chooses between CPU and CPU_GPU, and ML Kit falls back to CPU when no preferred configuration works.

Google asks for input images of at least 480x360 pixels and says the person should be represented by at least 256x256 pixels for best performance. For real-time use it recommends the base SDK with STREAM_MODE.

The limits Google states#

• One person. "Pose detection can only detect one person in an image." With two people, the landmarks go to the person detected with the highest confidence.
• The face must be visible. "The user's face must be present in order to detect a pose."
• Beta. The API "is offered in beta, and is not subject to any SLA or deprecation policy", and Google warns that changes may break backward compatibility.

The ML Kit Terms of Service, last modified May 14, 2025, say processing of the input data "fully happens on-device", and that the APIs "send metrics about the performance and utilization of the APIs in your app to Google". For more than one person in frame, start from multi-person pose tracking.

When it fits#

Our reading: ML Kit suits an Android and iOS app that wants one person's skeleton without shipping and versioning a model file, and that can accept beta terms. If you need the model file itself or an open-source licence, MediaPipe Pose Landmarker outputs 33 landmarks at the same indices; its variants are compared in MediaPipe Pose Landmarker models.