Preprocessing runs pose and equipment detection on a video ahead of time and saves the results on your device. When you later open that video for analysis, the saved data loads instantly and plays back perfectly smoothly—ideal for longer clips, slower phones, or when you want to use the most accurate (Heavy) model without worrying about real-time speed.
Steps
1. Open Video Preprocessing
Open the Video Preprocessing screen. Videos you have already processed appear at the top with badges for each model that was run, plus resolution, duration, and frame count.
2. Select a video
Tap Select Video and pick any clip from your device.
3. Choose which models to run
Enable any combination of Pose Detection, Barbell, Kettlebell, and Dumbbell detection. At least one model must be selected to start.
4. Set the pose detector and scale
If Pose Detection is on, pick a detector family (MediaPipe Heavy is recommended for preprocessing because accuracy matters more than speed when it runs offline) and a Processing Scale from 0.5x (faster) to 1.0x (most accurate).
5. Pick the processing hardware
Choose CPU, which works on every phone, or GPU, which is faster where supported. The app probes your device and tells you whether GPU is available.
6. Start preprocessing
Tap Start Preprocessing. A progress overlay shows overall and per-model progress with frame counts, keeps the screen awake, and lets you pause or cancel. It runs in a foreground service so it keeps going while the screen is on.
7. Use the saved data later
Next time you open that video for analysis, the app offers a Data Source choice: Use Saved Data, which loads instantly, or Analyze Live, which processes each frame in real time.
Tips
- Heavy model at 1.0x scale gives the most accurate results; because preprocessing is offline, slower processing does not affect playback later.
- Preprocessed videos are stored on your device. Delete any entry from the list to free up space.
- If preprocessing cannot finish, an error banner lets you retry without re-selecting your video and models.