Objectron 3D tfjs models: chair and cup are not correct

[luannd]

1. OS you are using: Mac OS

2. OS Architecture: Mac OS

3. Version of OpenVINO: same as your Docker

4. Version of TensorFlow: same as your Docker

5. Version of TensorRT: same as your Docker

6. Version of TFJS: 2.7.0

7. Version of coremltools: same as your Docker

8. Version of ONNX: same as your Docker

9. Download URL for .tflite IR model: directly from Mediapipe repo: https://github.com/google/mediapipe/tree/master/mediapipe/models

10. URL of the repository from which the transformed model was taken: use your Docker command

11. URL or source code for simple inference testing code

let image_tf = tf.browser.fromPixels(inputCanvasElement).asType('float32');  
let image_tf_resized = tf.image.resizeBilinear(image_tf, [224, 224]).expandDims(0).div(255.0);
prediction = model.predict(image_tf_resized);

[prediction_keypoints] = prediction[0].arraySync();
[prediction_prob] = prediction[1].arraySync();

12. Issue Details

I can convert the tflite models to tfjs but they seem to predict wrong results while testing by tfjs. The 9 points are kind of random and cannot be detected [correctly.] My preprocessing is the same as https://github.com/sfsrd/objectron. Not sure what's wrong in the conversion pipeline.

[PINTO0309]

Here are the inference results in tflite, converted by my tool. There doesn't seem to be any particular problem, so if there is a problem with the inference results when converting to tfjs, it may be a bug in tfjsconverter.

https://user-images.githubusercontent.com/33194443/109388909-4874f800-794d-11eb-8760-6c263b913935.mp4

[PINTO0309]

This is the model of CUP that I converted. saved_model.zip

https://youtu.be/LHCazNkr8uk

[PINTO0309]

Closed due to lack of progress.