MLKit's on-device Label(Image) Detection is very poor when compared to TensorFlow Label Detection

[Raju-567]

MLKit's on-device Label Detection is very poor when compared to TensorFlow Label detection. MLKit on-device model is not able to predict the images accurately .

When I debugged the both apps, I found that MLKit uses preview ImageFormat as NV21 and TensorFlow is using YUV_420_888 Image format.

ML kit android app
https://github.com/firebase/quickstart-android

TensorFlow android app
https://github.com/tensorflow/examples/tree/master/lite/examples/image_classification/android

And I even tried using tflite model from TensorFlow app as a custom model in the MLKit android app and I see the predictions results are not satisfactory (the results are same as ML Kits on device
Label Detection)

You can see the prediction results of MLKit and TensorFlow apps on water bottle and Laptop input Images.