Every question the Guides pages answer
The 11 guides pages answer 55 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.
Camera Pose Tracking for Fitness Apps: A Practical Guide
- What is the difference between the 33 landmarks in MediaPipe and the 17 in MoveNet?MediaPipe Pose Landmarker returns 33 landmarks in the BlazePose topology, a superset that adds face, hand, and foot points plus a per-point…
- Which landmark indices do I use for exercises?In MediaPipe's 33-point model the exercise joints are shoulders 11 and 12, elbows 13 and 14, wrists 15 and 16, hips 23 and 24, knees 25 and…
- Does camera pose tracking run on-device or send video to a server?Both MediaPipe Pose Landmarker and MoveNet run entirely on-device. Once the model files are loaded and cached, detection needs no network…
- Why do the left and right landmarks look swapped on screen?They are not swapped. Both models label joints from the subject's anatomical perspective, not the viewer's, so when someone faces the…
- Does pose estimation count reps and check form by itself?No. Pose estimation only returns joint coordinates, one set per frame. Rep counting is typically a threshold state machine over a joint…
How to Add Rep Counting to Your Fitness App
- Do I need machine learning to count reps from pose keypoints?No. In the common case rep counting is a threshold state machine over one scalar signal — usually a joint angle — not a trained model.…
- Why use two thresholds instead of one for counting reps?A single threshold lets sensor noise chatter across the line and fire many counts per real rep. Two separated thresholds create a…
- Which joint angle should I track for each exercise?Track the joint that bends most between the two ends of the movement. For a bicep curl that is the elbow angle from shoulder, elbow, and…
- Why is my rep count jumping or double-counting?The usual causes are unsmoothed jitter, thresholds set too close together, or occluded low-confidence landmarks spiking the angle. Add EMA…
- Does this work with both MediaPipe and MoveNet?Yes. The angle math and state machine only need x and y plus a per-keypoint confidence value, which both provide (visibility on MediaPipe,…
How to Add Real-Time Form Feedback to a Workout App
- How do I check squat depth from pose keypoints?Compute the knee angle at the vertex knee for the hip-knee-ankle joints using the atan2 interior-angle formula, and track the smallest…
- How do I detect a knee caving inward (valgus)?From a front view, compare the horizontal x-position of the knee to the ankle: when the knee drifts toward the body midline relative to the…
- When should form be evaluated during a rep?Evaluate at the rep transition, using the extreme values captured across the whole rep window rather than one frame. This ties form…
- Why show only one cue at a time?Firing several corrections at once is unusable for someone mid-workout. Pick the single highest-priority failed rule, for example depth…
- Is camera-based form feedback medical or injury-prevention advice?No. This technique only flags when a joint angle or alignment is outside a target range you define; it does not diagnose anything or…
How to Track Workouts Without a Wearable (Camera-Only)
- Can you really track a workout without a wearable?Yes, for the movement side of a workout. A phone camera plus an on-device pose model measures reps, range of motion, tempo, and form…
- What can a camera measure that a wearable cannot?A camera sees geometry over time, so it can measure reps, joint angles and range of motion, movement tempo, and form and alignment. Most…
- Can a camera measure calories burned?Only as an estimate. A camera can estimate calorie burn from movement, exercise type, duration, and a user's height and weight, but it is a…
- Do I need a paid SDK, or can I build camera tracking myself?You can build it on free, permissively licensed pose primitives like MediaPipe Pose Landmarker or TensorFlow MoveNet and write the rep and…
- Does camera tracking send my video to the cloud?It does not have to. MediaPipe Pose Landmarker and MoveNet both run inference on-device, so after the model loads no frames need to leave…
How to Add AI Workout Tracking to an iOS App (Swift)
- Does AI workout tracking with Vision work offline?Yes. Vision's body-pose detection runs entirely on-device, so no image data leaves the phone and no network connection is required. This is…
- Which iOS version do I need?The 2D body-pose request (VNDetectHumanBodyPoseRequest) requires iOS 14.0 or later. The 3D request (VNDetectHumanBodyPose3DRequest)…
- How do I count reps from body joints?Rep counting is a threshold state machine, not machine learning. Compute a driving joint angle, smooth it, and use two separate thresholds…
- Why do my skeleton overlays appear upside-down?Vision returns normalized coordinates from 0 to 1 with the origin at the bottom-left, while UIKit and Core Graphics use a top-left origin.…
- Should I use Apple Vision or MediaPipe on iOS?Vision ships in the OS with no dependencies and gives 19 joints, which is enough for angles and rep counting. MediaPipe Tasks for iOS gives…
How to Add AI Workout Tracking to an Android App (Kotlin)
- MediaPipe PoseLandmarker or ML Kit Pose Detection for Android?Both are from Google, both derive from BlazePose, and both emit the same 33 landmarks. ML Kit is a shorter one-line drop-in for Android and…
- Does the pose detection run offline?Yes. Both MediaPipe PoseLandmarker and ML Kit run entirely on-device, so after the model is bundled or downloaded no image data leaves the…
- How do I count reps without training a machine learning model?Rep counting is usually a threshold state machine, not machine learning. You compute a driving joint angle, smooth it, and toggle between a…
- Why do my rep counts jump or double-count?Almost always because of a single threshold or unsmoothed noise. Use two separated thresholds so the angle must fully commit to the other…
- Which camera angle and setup give the most accurate tracking?Use a side view for sagittal exercises like squats, curls and lunges so the joint angles sit in the image plane, and a front view for…
How to Add AI Workout Tracking to a React Native App
- Which frame-processor API does VisionCamera use in React Native?It depends on your installed version. The frame-processor API has changed across major versions: v3 and v4 used a useFrameProcessor hook…
- Should I use MoveNet TFLite or an ML Kit plugin for pose detection?Both are valid. Running a MoveNet or BlazePose .tflite model with react-native-fast-tflite inside the worklet gives you control over the…
- How do I count a rep from pose keypoints?Compute one driving joint angle (elbow for a curl, knee for a squat), smooth it, and run a two-state machine with two thresholds. The angle…
- Does this work in Expo Go, or do I need a development build?Expo Go alone will not work. Frame processors need native modules, and Expo Go ships a fixed set of native code you cannot add to, so…
- Can I skip building this and use a fitness SDK instead?Yes. Commercial fitness SDKs such as KinesteX, Kemtai, and QuickPose bundle the camera, pose model, smoothing, rep counting, and a prebuilt…
How to Add AI Workout Tracking to a Flutter App
- Do I need a paid API or the cloud for AI workout tracking in Flutter?No. google_mlkit_pose_detection runs the pose model on-device for free, so tracking works offline with no per-call cost. A paid commercial…
- How does the app count reps if pose detection is not counting them?Rep counting is separate logic you add on top of the landmarks. You compute a joint angle from three landmarks each frame and run a…
- Why use atan2 instead of acos for the joint angle?The atan2 of the cross and dot products is numerically robust and returns 0 to 180 degrees cleanly. The plain acos form can lose precision…
- Which package versions should I use as of 2026?As of 2026 google_mlkit_pose_detection is around 0.15.0, but versions move quickly, so check the current release on pub.dev and the package…
- Will this work on a phone's front camera and in a simulator?It works on both front and rear cameras, though the rear camera usually gives higher-quality landmarks. It needs a real device with a live…
How to Add AI Workout Tracking to a Web App (JavaScript)
- Which pose model should I use for web workout tracking, MediaPipe or MoveNet?MediaPipe Pose Landmarker returns 33 landmarks with face, hand, and foot detail plus a z depth value, which is useful for richer form…
- Does the camera video get sent to a server?No. MediaPipe Pose Landmarker and MoveNet both run inference locally in the browser using WASM, WebGL, or WebGPU. Once the model and WASM…
- Why does the browser never show a camera permission prompt?Almost always because the page is not on a secure origin. Browsers block getUserMedia outside HTTPS or localhost, and the call rejects…
- Why is my joint angle jumpy or wrong?A 2D angle only matches the true joint angle when the limb is roughly parallel to the camera plane, so use a side view for squats, curls,…
- Do I need a build step, or can I use a CDN?Both work. As of 2026 you can npm install @mediapipe/tasks-vision and bundle it, or import it and the WASM fileset from a CDN such as…
How to Improve Pose-Detection Accuracy in Your App
- Why are my pose landmarks jittery even when they are in the right place?That is high-frequency noise, not a placement error, and the fix is temporal smoothing. Apply a One-Euro filter to the raw landmark…
- What is the difference between the One-Euro filter and EMA for pose smoothing?An EMA blends each new value with the running average at a fixed rate, which is cheap and good for scalar signals like a joint angle or a…
- Should I use a bigger model to get better accuracy?A heavier tier (MediaPipe heavy, or MoveNet Thunder at 256x256 versus Lightning at 192x192) is more accurate but slower, so use it only…
- What confidence threshold should I use to reject bad landmarks?A common gate is around 0.5 to 0.7 on each landmark's visibility (MediaPipe) or score (MoveNet), but tune it per camera and gate harder on…
- Does the camera angle affect accuracy or just the view?It affects your measurements directly. A 2D landmark angle only equals the true joint angle when the limb moves roughly parallel to the…
How to Evaluate AI Motion Tracking SDKs
- Why not just compare the vendors' published accuracy figures?Because no public benchmark exists for this category, so there is nothing to compare. Every figure published by a camera-coaching vendor…
- How long does a motion SDK bake-off take?Budget two weeks of focused effort: roughly one week filming and labelling the corpus, and one week integrating each SDK and running it.…
- What should a motion tracking vendor put in writing before you sign?Six things a test corpus cannot measure: current pricing and the unit it scales on (per active user, per session, per device), on-device or…
- Should a shipped exercise library count toward the score?No. Treat it as a fit question that filters your shortlist, not a scored criterion that ranks it. Whether a bundled content library helps…
- What if an SDK will not let me feed it recorded clips?Record the refusal as a finding and score it, then fall back to a uniform substitute: play each clip full-screen on a second display in…
Back to the Guides hub, or see every question this site answers.