Integrating Custom TensorFlow Lite Model in Flutter App: Object Detection Not Working After Replacement

Question

I'm attempting to integrate a custom TensorFlow Lite (TFLite) model, created using Teachable Machine, into a Flutter app. Initially, I followed the live object detection example provided in the tflite_flutter plugin repository, specifically the live_object_detection_ssd_mobilenet example.

I successfully ran the example with the provided model (ssd_mobilenet.tflite), but when I replaced it with my custom TFLite model, the app ceased to function properly. Specifically, it no longer displayed bounding boxes and statistics for detected objects.

Here are the steps I took:

Downloaded and replaced the default ssd_mobilenet.tflite model with my custom model.
Updated the model path in detection_service.dart to point to my custom model.
I check Netron and ensured that the mlModelInputSize in the code matches the input size used during training (224 instead of the default 300).

class _DetectorServer {
  /// Input size of image (height = width = 300)
  static const int mlModelInputSize = 300;

  /// Result confidence threshold
  static const double confidence = 0.5;
  Interpreter? _interpreter;
  List<String>? _labels;

  _DetectorServer(this._sendPort);

  final SendPort _sendPort;

  ...
}

Additional:

I'm new to tflite. I need a clear guidance to use my own model in the example that I provided above. If you have another suggestion or knowledge related to this issue, please let me know. Thank you.

Chathura Chamikara · Accepted Answer · 2024-05-12 11:09:51Z

0

First, your model is a classification model, not an object detection model. So, there are no bounding boxes. In Netron, you can see your output shape is [1,29], meaning there are 29 classes. If you print the output just after you infer the model, you will get the probabilities for each class like [[0.002,0.004,0.5,....]].

If you need to train your custom model for object detection, you will need to train it with an annotated dataset with bounding boxes. You can search for a Google Colab notebook to train your annotated data with SSD mobile net or something like yolo. The teachable machine is just for classification models.

answered May 12, 2024 at 11:09

Chathura Chamikara

1246 bronze badges

Sign up to request clarification or add additional context in comments.

3 Comments

zimmer Over a year ago

I tried to use my model and run another code example for image classification from this repo: github.com/tensorflow/flutter-tflite/tree/main/example/…, but the code is not working. is there a possible way to use teachable machine model to that example? Most tutorial I found is outdated and not using tflite_flutter package

Chathura Chamikara Over a year ago

You are using the correct example that suitable for your model. Without any error logs, or code samples, I would not be able to help you.

zimmer Over a year ago

I post the issue on github, I appreciate it if you can take a look at it. github.com/tensorflow/flutter-tflite/issues/216

Collectives™ on Stack Overflow

Integrating Custom TensorFlow Lite Model in Flutter App: Object Detection Not Working After Replacement

1 Answer 1

3 Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

3 Comments

Your Answer

Sign up or log in

Post as a guest

Related