Customized tflike model can't work in iphone app

[gzhhong]

Have I written custom code (as opposed to using a stock example script provided in MediaPipe)

Yes

OS Platform and Distribution

iOS

Python Version

3.9

MediaPipe Model Maker version

No response

Task name (e.g. Image classification, Gesture recognition etc.)

hand landmarks

Describe the actual behavior

Customized tflite model doesn't work in iPhone app

Describe the expected behaviour

The tflite model with the same meta data should work in iphone App as the old one

Standalone code/steps you may have used to try to get what you need

https://github.com/gzhhong/mediapipe-task

Other info / Complete Logs

The readme describe how to reproduce the issue. I use the formal iphone app example at https://github.com/google-ai-edge/mediapipe-samples/tree/main/examples/hand_landmarker/ios

[gzhhong]

Hello, just add more findings here. The output of keras suppose to be landmarks, handedness, presence_score, world_landmarks, but the output of tflite may be in another order, and that seems cause the problems. I think the key is to make the order of output of tflite is same as the output of keras, and the order should be landmarks, handedness, presence_score, world_landmarks.

[gzhhong]

Hello, I want to train a new tflite model and replace the one in hand_landmarker.task in the application https://github.com/google-ai-edge/mediapipe-samples/tree/main/examples/hand_landmarker/ios. Now my tflite model will ouput tensors in shape [1, 63], [1, 1], [1, 1], [1, 63], which represent the landmarks, handedness, presence_score and world_landmarks. The coordinates of landmarks are the pixel coordinates in the image of 224*224. Now the issue is, when the landmarks displayed in the iphone application, the landmarks just cover a small part of the hand image, it seems the coordinates not transferred from 224 to the screen. The landmarks sometimes shrink as one point.