Therefore my first mobile applications:
1. Snapshot classification in mobile phone.
Classification takes ~ 120 - 130 ms.
2. Stream classification in mobile phone.
Classification takes ~ 80 - 90 ms.
The applications are tested on Xiaomi Mi 9 mobile phone (Snapdragon 855, 6GB, Android 9, Chrome). Predicted probabilities are very noisy. Therefore see Stream classification test where classified images are plotted to the second canvas. On my HP laptop default video stream (640x480) is very noisy too. Remote controls are used for classification tests.
The application uses MobileNet_v2_1_224 TFjs model (32 bit weights = 14 MB). "float16 data" are slightly (~10%) faster. Unfortunately, I can't use quantized weights yet. Front and Rear Camera Access with JavaScript's getUserMedia() is used as prototype. Sorry for minimalistic design.