Every question the AI Motion pages answer

The 28 ai motion pages answer 121 named questions between them. Every one is listed below, and every link lands on the paragraph that answers it rather than the top of the page.

Pose Estimation Models Compared: MediaPipe, MoveNet, YOLO and More

How many keypoints does each model detect?MediaPipe/BlazePose predicts 33 body landmarks, a superset of the 17 COCO keypoints that adds face, hand, and foot points. MoveNet,…
Which model should I use for a single-user fitness app?For a single user on-device, MediaPipe/BlazePose is the common default: its 33 landmarks and monocular 3D world landmarks suit joint-angle…
Are these pose models free to use commercially?Not all of them. MediaPipe/BlazePose, MoveNet, and PoseNet are Apache-2.0, the most permissive here, though the MoveNet MultiPose model…
Which of these models output 3D landmarks?Among the common on-device options, only MediaPipe/BlazePose: alongside its 2D image-coordinate landmarks it returns 3D world landmarks in…
Which models handle multiple people at once?YOLO-pose detects multiple people and their keypoints in one pass, and OpenPose is a bottom-up multi-person system. MediaPipe/BlazePose is…
Should I pick MoveNet or MediaPipe for a fitness app?Our recommendation: choose by what you compute downstream. MediaPipe's 33 landmarks with 3D world landmarks favor form feedback where joint…

2D vs 3D Pose Estimation for Fitness Apps

Do I need a depth camera or LiDAR for 3D pose?No. A normal RGB phone camera is enough for both 2D pose and monocular 3D world landmarks — models like BlazePose infer depth from a single…
How reliable is monocular 3D depth?Less reliable than the x and y axes, and less reliable than multi-camera or depth-sensor 3D. Predicting true depth from one image is…
When is 2D enough and when do I need 3D?2D is enough when the joint angle you measure stays in the camera plane — for example an elbow angle filmed side-on or a knee angle filmed…
Is 3D form feedback accurate enough for injury or medical advice?No. Camera-based form feedback, whether 2D or monocular 3D, is a coaching aid, not medical or physical-therapy advice. Monocular depth…

Pose Estimation Accuracy: What It Means and What Drives It

What do PCK, OKS, and MPJPE measure?PCK (Percentage of Correct Keypoints) is the fraction of 2D keypoints within a distance threshold of ground truth, with the PCKh variant…
What degrades pose estimation accuracy the most?Low or backlit lighting, occlusion of joints behind the body or objects, extreme camera angles, too much distance from the camera, loose or…
How do confidence and visibility scores help?Pose models emit a per-keypoint confidence or visibility score (some, like BlazePose, add a presence score) that indicates how sure the…
What is the One-Euro filter and why is it used?The One-Euro filter is a low-pass filter with an adaptive cutoff used to reduce frame-to-frame jitter in real-time pose tracking. It…
How do I evaluate accuracy for my own app?Do not trust a leaderboard score alone. Record footage that matches production (your camera, your users' lighting, distances, angles,…

Multi-Person Pose Tracking: Top-Down vs Bottom-Up

What is the difference between top-down and bottom-up pose estimation?Top-down first detects each person, then runs a pose model on each cropped box, so per-person accuracy is generally higher but inference…
Does my fitness app need multi-person pose tracking?Usually not. Most fitness apps are single-user, with one person in front of one camera, and single-person models such as…
Which models support multi-person pose?OpenPose is the classic bottom-up multi-person system using Part Affinity Fields to group keypoints. YOLO-pose (Ultralytics) is…
How are people tracked across frames?Per-frame pose estimation does not by itself keep identities stable, so multi-person work adds a separate tracking layer that associates…
Are there licensing concerns with multi-person models?Yes, and they are a real decision point. OpenPose ships under a non-commercial/academic research license, so commercial use requires a…

On-Device vs Cloud Pose Estimation: Which to Choose

Is on-device or cloud pose estimation better for a fitness app?For most consumer fitness apps, on-device is the better default. Running the model on the phone with a toolkit like ML Kit Pose,…
Why is streaming workout video to a server a privacy concern?A workout video is raw footage of a person's body, face, and home - sensitive personal and biometric data, not a neutral rep count. When…
Can I keep video private but still use a server?Yes, with a hybrid architecture. Run pose detection on the edge device, then transmit only the text-based landmark coordinates - not the…
Does cloud pose estimation cost more than on-device?Generally yes, in a recurring way. On-device inference runs on hardware the user already owns, so there is no per-frame server bill. Cloud…
Does on-device pose estimation work on older phones?It can, but performance depends on the device. Phones with a strong GPU or neural accelerator run pose models smoothly, while older or…

Real-Time Pose Estimation: Frame Rate, Latency, and Model Trade-offs

What frame rate do I need for real-time pose estimation?Smooth, interactive tracking is commonly framed around 30 fps, but usable coaching can work at lower rates, so treat any single number as a…
Should I use MoveNet Lightning or Thunder?Lightning is tuned for speed with a smaller model and lower input resolution, so it suits live rep counting, real-time overlays, and…
What should I do if inference can't keep up with the camera?Drop frames rather than queue them. If you queue frames the model can't process in time, feedback drifts further and further behind the…
Why is temporal smoothing needed?Per-frame pose estimation jitters, so the keypoints wobble even when the subject holds still, and raw keypoints are too noisy to derive…
Does continuous pose estimation drain the battery?Yes. A continuous camera plus continuous inference is a sustained load that drains battery and heats the device over a full workout, and a…

Pose Estimation Hardware Requirements: What You Actually Need

Do I need a depth camera or LiDAR for pose estimation?No. A normal RGB smartphone camera is enough for 2D pose and for monocular (single-camera) 3D pose - both run from a single RGB frame.…
Does pose estimation need a GPU or a special AI chip?It runs faster and uses less battery with one, but it is not strictly required. Hardware acceleration - a mobile GPU, or an NPU/Neural…
Will pose estimation work on older or low-end phones?Often yes, but expect lower frame rates. Older and budget devices lack an NPU and have weaker GPUs and CPUs, so heavier models can drop…
What hardware do I need to run pose estimation in the cloud?Server-side pose inference typically needs GPU hardware to run heavier models at frame rate, which is a recurring per-frame or per-minute…
Do lighting and camera position affect the hardware I need?They affect accuracy more than they change the hardware, but they matter a lot. Accuracy degrades in low light, backlight, and harsh…

How Rep Counting Works: The Algorithm Explained

What signal does a rep counter actually track?It reduces the pose to one number that oscillates once per rep. Most commonly that is a joint angle from three keypoints (elbow angle for a…
What is hysteresis and why does it stop double-counting?Hysteresis means using two different thresholds with a gap between them instead of one. To count a rep you must cross a low threshold into…
How do you handle noise and false reps?Smooth the keypoints or the derived angle before any detection (a One Euro filter is a common low-cost choice), use hysteresis rather than…
Do I need a depth camera to count reps?No. A normal RGB phone camera is enough. Rep counting tracks 2D joint angles that stay in the camera plane, so no depth sensor or LiDAR is…
How accurate is camera-based rep counting?It varies by exercise, dataset, and the user's form, so there is no single number to quote. Partial reps, very fast reps, occlusion, and…

How Camera-Based Form Feedback Works

How does a form checker actually measure my form?It takes three keypoints that meet at a joint (for a squat, hip, knee, and ankle), computes the angle between the two limb segments, and…
Do I need a depth camera or special sensor for form feedback?No. A normal RGB phone camera is enough to track keypoints and compute joint angles. No LiDAR or depth sensor is required. The catch is…
Can I use camera form feedback for injury rehab or physical therapy?Treat it only as general coaching, never as medical or physical-therapy advice. Monocular pose estimation is not accurate enough for…
Why does form feedback get worse from some camera angles?Any joint angle that projects toward or away from the camera falls into the depth axis, which is exactly where a single-camera estimate is…
How accurate is camera-based form feedback?Good enough for gross-movement coaching, not for degree-level or clinical claims. Accuracy depends on lighting, clothing, camera angle,…

Build vs Buy: AI Motion Tracking

Isn't the hard part just picking a pose model?No. Choosing MediaPipe, MoveNet, or Apple Vision is the easy first step. The real, ongoing work is everything after the keypoints arrive:…
When does building your own pipeline make sense?Build when camera-based motion tracking is your core differentiator, you have computer-vision/ML engineers, you need offline use and zero…
What does buying a motion-tracking SDK cost?A recurring per-user or per-device fee, plus less control over the underlying model and its roadmap. Free tiers and paid plans both exist,…
Can I mix building and buying?Yes, and most teams do. The common hybrid is to adopt an on-device pose model or SDK for detection, then build your own rep, form, and…
Do I need a depth camera either way?No. A normal RGB phone camera is enough for both 2D and monocular 3D pose, whether you build or buy. Monocular 3D depth is estimated rather…

MediaPipe Pose Landmarker Models: Lite vs Full vs Heavy

Can I swap Pose Landmarker variants without changing code?Mostly yes, and that is the point of the family. All three variants emit the same 33 landmarks with the same per-landmark values (x, y, z,…
What does the /latest/ segment in the Pose Landmarker model URL mean?It is the version slot in the download path on Google's mediapipe-models storage bucket - a pointer to the newest published build rather…
Why is the pose_landmarker .task file bigger than the model size in the model card?The model card's figures - 3 MB lite, 6 MB full, 26 MB heavy - describe the landmark models alone, while the .task downloads are bundles…
What does float16 mean in the Pose Landmarker download path?It is the precision segment of the path - it names the numeric precision of the bundled model weights, in this case half-precision floating…
Does pose_landmarker_heavy add extra landmarks or multi-person tracking?No. Heavy outputs the same 33 landmarks as lite and full, and the model card states the model tracks only one person on scene if multiple…

VNDetectHumanBodyPoseRequest: Apple Vision Body Pose on iOS

Why do Apple's 2D and 3D body pose requests return different joint counts?They use different skeletons. VNDetectHumanBodyPoseRequest exposes 19 named joint constants including face detail (nose, eyes, ears) plus…
Does Apple publish a model card or accuracy numbers for Vision body pose?No. Apple's documentation is an API reference: it contains no model card, no accuracy figures, no evaluation dataset, no…
What OS versions do the Vision body pose requests require?Apple documents VNDetectHumanBodyPoseRequest (2D) as available from iOS 14.0, iPadOS 14.0, macOS 11.0, Mac Catalyst 14.0, tvOS 14.0, and…
Can Vision run body pose on recorded video instead of a live camera?Yes. VNVideoProcessor is Apple's documented object for offline analysis of video content: create it with init(url:) pointing at a video…
Should a cross-platform fitness app use Apple Vision for pose estimation?Usually not as its only pose layer. Vision runs on Apple platforms only, so an Android sibling app needs a second implementation with a…

Pose Estimation Model Licenses: Which Can You Ship Commercially?

Is MoveNet licensed under Apache 2.0?For the single-person models, yes: the MoveNet.SinglePose model card, which covers Lightning and Thunder, says "Licensed Under Apache…
Is MediaPipe free for commercial use?In our reading, yes, with conditions. The google-ai-edge/mediapipe LICENSE file and the BlazePose GHUM 3D model card both state Apache…
Does the AGPL apply if YOLO pose only runs on my server?In our reading, server-only use does not by itself put Ultralytics YOLO outside AGPL-3.0. Section 13 of the license says a modified version…
Can I use OpenPose in a commercial fitness app?Not under the license in its repository. CMU's LICENSE is headed "ACADEMIC OR NON-PROFIT ORGANIZATION NONCOMMERCIAL RESEARCH USE ONLY" and…
How do ML Kit's pose detection terms differ from MediaPipe's license?MediaPipe gives you Apache-2.0 code and a model card you can read, while ML Kit Pose Detection is a closed SDK under the ML Kit Terms of…

MoveNet Lightning, Thunder and MultiPose: Inputs, Outputs and Licence

What input size do MoveNet Lightning and Thunder expect?Google's MoveNet.SinglePose card gives an int32 tensor of 192x192x3 for Lightning and 256x256x3 for Thunder, RGB with values from 0 to 255,…
In what order does MoveNet's output tensor store coordinates?y first, then x, then the score. The SinglePose card describes a 1x1x17x3 float32 tensor whose last dimension holds the yx coordinates…
How many people can MoveNet MultiPose track?Up to 6. The MoveNet README describes MULTIPOSE_LIGHTNING as a multi-pose detector that detects up to 6 poses, the MultiPose card says the…
Where does the TF.js pose-detection package load MoveNet weights from?From tfhub.dev by default: the MoveNet constants file points at movenet/singlepose/lightning/4, movenet/singlepose/thunder/4 and…
What confidence threshold does Google suggest for MoveNet keypoints?0.3. The MoveNet.SinglePose card says the model predicts all 17 keypoints even when they are occluded, gives keypoints outside the frame…

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

What is the difference between ML Kit's base and accurate pose detection SDKs?Speed against coordinate accuracy, and size. Google's Android guide lists pose-detection at about 10.1 MB and about 30 FPS on a Pixel 3XL,…
Can ML Kit pose detection track more than one person?No. Google's overview says pose detection can only detect one person in an image, and that with two people in frame the model assigns…
Is the Z coordinate from ML Kit pose detection real depth?No. Google calls it an experimental value measured in image pixels like X and Y that is not a true 3D value, with its origin approximately…
What image resolution does ML Kit pose detection need?Google's Android guide asks for images of at least 480x360 pixels and says the person should be represented by at least 256x256 pixels for…
Why does ML Kit return pose landmarks outside the image?Because it always returns all 33. When the person is not completely inside the image, Google's Android guide says the model assigns the…

Ultralytics YOLO Pose Models: Sizes, Keypoint Output and Licence

What do the n, s, m, l and x YOLO pose model sizes trade off?Accuracy against speed and size, on Ultralytics' own benchmark. Its table runs from YOLO26n-pose at 57.2 mAP pose 50-95, 40.3 ms on CPU…
Does Ultralytics YOLO pose return keypoints for every person in the image?Yes, one set per detected instance. Ultralytics documents result.keypoints with the shape (N, K, 2 or 3), N instances of K keypoints,…
What is the third value in a YOLO pose keypoint?Visibility or confidence, where the model provides it. coco-pose.yaml declares kpt_shape as 17 keypoints by 3 dimensions and explains the…
Can a YOLO pose model be exported to Core ML or TensorFlow Lite?Yes, according to Ultralytics' export table: format coreml writes a .mlpackage and format litert writes a .tflite file, alongside ONNX,…

RTMPose (MMPose): Model Sizes, Keypoint Sets and Deployment

Which keypoint sets do RTMPose models come in?Three body sets, per the RTMPose README: 17 keypoints defined as COCO, 26 keypoints defined as Halpe26, and 133 keypoints defined as…
Does RTMPose need a separate person detector?In the pipelines its README documents, yes. The pipeline performance table pairs each RTMPose model with an RTMDet person detector, and the…
Can RTMPose run without PyTorch?Yes. The RTMPose README lists rtmlib, which runs RTMPose, DWPose, RTMO and RTMW models with OpenCV, ONNX Runtime or OpenVINO and does not…
How fast is RTMPose on a mobile chip?The RTMPose README reports RTMPose-s at 72.2% AP on COCO with 70+ FPS on a Snapdragon 865 chip, measured by the project itself after…

OpenPose Models: BODY_25 vs COCO, Output Format and Licence

What is the difference between OpenPose's BODY_25 and COCO body models?Seven points and the numbering. OpenPose's poseParameters.cpp lists 25 body parts for BODY_25 and 18 for COCO; BODY_25 has every COCO name…
How is OpenPose's JSON keypoint output laid out?One file per frame with a people array, per OpenPose's output document. Each person's pose_keypoints_2d is a flat list of x, y and…
Does OpenPose need an NVIDIA GPU?No. The OpenPose README lists CUDA builds for Nvidia GPUs, OpenCL builds for AMD GPUs and non-GPU, CPU-only builds, on Ubuntu, Windows, Mac…
Can OpenPose estimate 3D poses?Only in a multi-camera setup. The OpenPose README documents 3D real-time single-person keypoint detection by triangulation from multiple…

Squat Detection With Pose Estimation: Landmarks, Camera, Rep Logic

Where should the phone go to count squats?In Oliosi and colleagues' 2026 study, diagonal (45°) and front views beat side views at every distance, and the best squat result came from…
Which pose landmarks does a squat counter need?Hip, knee and ankle for the knee angle, and shoulder, hip and knee for the hip angle; every keypoint scheme on this site has all of them.…
What knee angle should count as the bottom of a squat?No study we read gives a universal number. Sousa and colleagues' example squat enters the descent below a knee angle of 100 and a hip angle…
How accurate is camera-based squat counting?It depends on the setup. One phone pipeline averaged a 61.5% detection rate across 12 phone positions and reached 95.5% from its best one;…

Push-Up Counter With Computer Vision: Landmarks, Camera, Rep Logic

Where should the camera go to count push-ups?In Oliosi and colleagues' 2026 study, diagonal views at 90 to 180 cm counted push-ups best and a frontal view at 360 cm worst, and the…
What elbow angle marks the top of a push-up?Lei and colleagues used 140 degrees as an empirical, adjustable elbow-angle threshold for the high position, with a maximum torso-to-ground…
Why does a push-up counter miss repetitions?The studies we read point to self-occlusion by the torso, floor-level poses that pose models handle badly, what one study logged as…
How accurate is camera-based push-up counting?One phone pipeline averaged a 61.1% detection rate across 12 phone positions and reached up to 85.7% from its best diagonal views, while a…

Lunge Detection With Pose Estimation: Landmarks, Camera, Rep Logic

Which camera angle works best for tracking lunges?In Baldinger and colleagues' 2025 study with four cameras between the frontal and sagittal views, back-viewing cameras performed best and…
Should a lunge counter track one knee or both?Both. Sousa and colleagues' lunge module requires both the front and rear knee angles to drop below a threshold to enter the lunge and to…
How accurate are 2D knee angles during a lunge?Not very, unless the view is controlled. Baldinger and colleagues measured OpenPose knee-angle RMSE of 15.50 to 29.44 degrees against…

Plank Detection With Pose Estimation: Landmarks, Camera, Hold Logic

How does a camera app time a plank?In Kim and colleagues' 2026 school study, a proper plank posture started a timer and an incorrect posture stopped it and triggered…
Which landmarks does a plank tracker need?Neither study we read names the landmarks of its plank rule. A horizontal-alignment check needs points along the body line, such as…
How accurate is camera-based plank detection?Sousa and colleagues report 100% for a class of nine planks, with a 95% confidence interval of 69.2 to 100%, and treat it as preliminary…

Deadlift Pose Estimation: Landmarks, Camera, Rep Logic

Can a 2D pose model see a rounded back during a deadlift?Not directly. Shoaib and colleagues state that projected 2D keypoints cannot recover depth-sensitive quantities such as spinal curvature or…
Which signal should a deadlift rep counter use?Sousa and colleagues' state machine uses a torso angle, forward-leaning at the start and extended in the lift, with a check that the elbows…
Where should the camera go for deadlift tracking?Mercadal-Baudart and colleagues placed phones about 3 m from the athlete, on the floor and at about 1.5 m height, and Shoaib and colleagues…
How accurate is camera-based deadlift analysis?Sousa and colleagues counted deadlifts with 95% accuracy. Shoaib and colleagues' form classifier reached 89.5% on new videos of its three…

Bicep Curl Counter With Pose Estimation: Landmarks, Camera, Rep Logic

Should a curl counter use the elbow angle or the hand position?Both appear in the studies. Chen and colleagues measured the angle at the elbow for curls, and Hwang and colleagues counted from the…
Where should the camera go to count bicep curls?Hwang and colleagues filmed from 1.5 m in front at 1.2 m height, Noteboom and colleagues from 1.5 m in front at 1.5 m height angled 15…
Why does a curl counter overcount or miss reps?In the studies we read, the equipment blocked the wrist, a second person in the background broke the skeleton, and short or uneven sets…
How accurate is camera-based bicep curl counting?Hwang and colleagues counted biceps curls with larger errors than presses and rows, with mean absolute errors across their five exercises…

Jumping Jack Counter With Pose Estimation: Landmarks, Camera, Rep Logic

What signal should a jumping jack counter use?Cho and colleagues defined a jumping jack by two cycles, the distance between the hands and the distance between the feet, each from its…
Where should the camera go to count jumping jacks?Both studies we read placed a Kinect depth camera in front of the person: on a tripod at 1.4 m height with a vertical angle of −8° and the…
How accurate is camera-based jumping jack counting?No study we read reports it. The two studies that measured jumping jacks with a camera used a depth sensor and analysed cycle timing in one…

Sit-Up Counter With Pose Estimation: Landmarks, Camera, Rep Logic

Which landmarks does a sit-up counter use?Lei and colleagues used the distance between elbow and knee for the sitting-up position and the angle between the body and the horizontal…
Where should the camera go to count sit-ups?To the side, on the evidence we read: Gu and colleagues' camera branch recognised sit-up frames 95.0% of the time from the side, 88.5% from…
How accurate is camera-based sit-up counting?No study we read reports it across a group of users. Gu and colleagues measured exercise recognition rather than counting, and Lei and…

Pull-Up Counter With Computer Vision: Landmarks, Camera, Rep Logic

How does a pull-up counter know where the bar is?In Lei and colleagues' 2025 system, it does not need to be told. Once the hands grip the bar, their height stays constant, so the system…
What counts as the top of a pull-up for a camera counter?Lei and colleagues count the top when the mouth rises above the estimated bar height, and Sousa and colleagues require the elbows to flex…
How accurate is camera-based pull-up counting?Lei and colleagues counted 96.6% of pull-ups correctly against human assessors, with lower accuracy for ordinary students and teachers than…

Overhead Press Pose Estimation: Landmarks, Camera, Rep Logic

Which angles does an overhead press counter use?Sousa and colleagues used the elbow and shoulder angles: below a threshold with the weights at shoulder level, above a higher threshold at…
Where should the camera go to track an overhead press?Hwang and colleagues filmed seated presses from 1.5 m in front at a height of 1.2 m, and Sousa and colleagues worked with videos from mixed…
How accurate is camera-based overhead press counting?Hwang and colleagues counted seated shoulder presses accurately, with small average absolute errors, and Sousa and colleagues counted…

Back to the AI Motion hub, or see every question this site answers.