An expandable section inside Video Settings for power users. These controls tune the detection engine itself—which models run, how they are accelerated, how confident they must be, and how many detector resources run in parallel. Changes that rebuild detector resources take effect when you tap Done.
Pose Detection
Choose pose scheduling, model routing, hardware acceleration, confidence behavior, and benchmarking.
Pose Round Robin and Pose Pool Size
Runs multiple pose-detector resources through a shared pool to increase throughput at the cost of memory. When enabled, Pose Pool Size selects two, three, or four detector resources. When several analysis modes run together, the smallest selected detector pool limits how many frames can be in flight.
Adaptive Pose Detector
Beta. Lets the app switch between the Lite, Full, and Heavy models per frame based on pose difficulty and your speed-to-accuracy preference.
Adaptive mode keeps Lite, Full, and Heavy loaded so it can switch without rebuilding detectors during playback. This uses more memory and can overwhelm memory-constrained phones. It is designed for 30 FPS video and should not be used together with Adaptive Scaling.
Detector Routing Preference
Beta. When the adaptive detector is on, a Performance-to-Accuracy slider shifts borderline frames toward faster or more accurate model choices without overriding obviously easy or hard poses.
Pose Model
Shows the model currently running and picks the fixed family used when Adaptive Pose Detector is off. Changing the model closes the current pose pool and builds the new family after you tap Done.
- Lite—lightweight MediaPipe model
- Full—balanced MediaPipe model
- Heavy—most precise MediaPipe model
Pose Delegate
Selects whether MediaPipe pose detection runs on the CPU or with GPU acceleration.
- CPU Only—standard processing
- GPU—hardware accelerated
Detection Confidence
How sure the app must be before it reports a person in the frame. Lower finds people more easily but with more mistakes; higher waits for a cleaner first lock.
Presence Confidence
How sure the app must be that the person is still there once a pose is found. Lower keeps a pose alive longer; higher drops shaky detections sooner.
Tracking Confidence
How easily the app keeps following the same pose from one frame to the next. Lower tracks through blur and fast motion; higher restarts sooner when tracking looks uncertain.
Object Detection
Configure independent Barbell, Kettlebell, and Dumbbell pools, then choose the delegate shared by all three bundled object detectors.
Equipment Round Robin and Pool Size
Each equipment detector has independent scheduling controls. Larger pools can improve throughput but use more memory.
- Barbell Round Robin enables a multi-resource barbell-detector pool; Barbell Pool Size chooses two, three, or four resources.
- Kettlebell Round Robin enables a multi-resource kettlebell-detector pool; Kettlebell Pool Size chooses two, three, or four resources.
- Dumbbell Round Robin enables a multi-resource dumbbell-detector pool; Dumbbell Pool Size chooses two, three, or four resources.
Object Detector Delegate
Chooses how the bundled object detectors run on your device.
- CPU Only—most compatible, runs on the main processor
- GPU Delegate—faster on supported devices
- NNAPI—routes through any device-specific neural acceleration drivers
Custom Tracker
Configure the separate pool and visual tracker model that power Custom Object Tracking. These trackers run through OpenCV, so the object detector delegate picker above does not apply to them.
Custom Tracker Round Robin and Custom Tracker Pool Size
Runs two, three, or four custom-tracker resources in a shared pool. Use Custom Tracker Pool Size to choose the count when round robin is enabled.
Tracker Model
Selects the tracker used when you follow a custom target. Changing this rebuilds the tracker and requires selecting the target again.
- NANO—lighter tracker for faster live playback on constrained devices
- VIT—heavier transformer tracker for tougher targets when frame time allows
Performance Benchmarks
Measure how each processing delegate performs on your device before committing to one.
Object Detector Benchmark
Times the bundled object detectors across the available delegates so you can compare CPU, GPU, and NNAPI speed on your own hardware.
Pose Model Benchmark
Times the pose models across delegates to help you pick the fastest reliable option for pose detection.
Recovery and Defaults
Reset to Recommended Settings
Restores the advanced settings recommended for the current device, including the fixed pose model, delegates, confidence thresholds, and detector-pool configuration.
Reset All Detectors
Closes and rebuilds every pose, equipment, and custom-tracker resource pool with the currently applied settings. Use it to recover the detection engine without changing your chosen configuration.