I'm used to working with MediaPipe in Python on a Windows 11 desktop setup. In this environment, if VIDEO mode is selected, MediaPipe automatically reads the video frame count, and the detected frames are processed directly. Recently, I started working on an Android implementation of my existing project using MediaPipe. However, I noticed that on Android (specifically for facial landmark extraction), MediaPipe requires a time interval between frames to be specified. Because of this, I end up with different frame counts and landmark predictions when processing the same video on both devices, even when using the same face_landmarker.task file. My main question is: Is there a way to disable the time interval specification on MediaPipe for Android, so that it instead reads the video file properties directly (like on desktop)? What I've tried so far: I attempted to match the expected frame interval more closely to the video properties read by OpenCV (e.g., 29 FPS vs 30 FPS). Despite this adjustment, I still get different predictions and frame counts. This isn't too surprising, since the FPS value I read isn't that close to a clean integer. Continue reading...
