How to improve the accuracy of Yolov4-tiny.tflite in flutter

Question

I have trained the yolov4-tiny file and successfully converted it to tflite. There were no problems during the computer test, but I encountered difficulties in installing it on the phone.

As I put my tflite into the flutter, I had different test results from my computer. And the result is not good, the accuracy is much lower, and the recognition error is often. I think the problem appear at classierYOLOv4, but i'm not sure where is wrong.

Is there any suggestion for me to fix this problem?

Thanks a lot for your help!

Here is computer test image:

Here is the code:https://github.com/piggychu0w0/flutter-yolov4tiny-tflite

Answer 1

I think the problem here is your input not being scaled to (0,1). You have to add a normalize layer into your model or scale your input first.

Edit: Using this code to normalize image

TensorImage getProcessedImage(TensorImage inputImage) {
    padSize = max(inputImage.height, inputImage.width);
    if (imageProcessor == null) {
      imageProcessor = ImageProcessorBuilder()
          .add(ResizeWithCropOrPadOp(padSize, padSize))
          .add(ResizeOp(INPUT_SIZE, INPUT_SIZE, ResizeMethod.BILINEAR))
          .add(DequantizeOp(0, 255.0))
          .build();
    }
    inputImage = imageProcessor.process(inputImage);
    return inputImage;
  }

and load the input image

TensorImage inputImage =  TensorImage(TfLiteType.float32);
inputImage.loadImage(image);
inputImage = getProcessedImage(inputImage);