I follow tutorial on 基于 MoveNet 和 TensorFlow Lite 的人体姿势分类 to create custom classification tflite model.
The accuracy of h5 and tflite model is about 90%.
After I copied classification.tflite and labels.txt to android project.
(replace old files in assets and edit download gradle)
The accuracy of classification on android app seem to be randomed.
What did I do wrong?
Hi @Mike_S.C ,
Hope you solved your issue by this time. If not resolved, please try the following pointers.
First check the availability of sufficient memory on your device. Insufficient memory leads to incorrect results.
Then ensure that the input pre-processing steps like (normalization, scaling and transformations, etc) applied to the data in the core model are exactly the same as those applied in the Android app before feeding it to the TFLite model. Slight differences in pre-processing may have significant effect on the model’s output.
Make sure the input data formats for both the model and android app are consistent.
After ensuring these changes please try inference with a single image.
If not resolved yet, please provide the details of your device configuration along the reproducible code.
Thank You