How to predict on one image?

Hello,

I’m still trying to reproduce this example using Keras.

I was able to train the model, but I have problems with predicting. I want to see a prediction for a one single image. Here’s my code:

    model.load_weights("camvid_model.h5")
    test_image = cv2.imread("Seq05VD_f01980.png") # It's an image from test folder of CamVid dataset
    ready_image = np.expand_dims(test_image, axis=0).astype('float')
    predicted_image = model.predict(ready_image)[0, :, :, :]
    predicted_image = (predicted_image * 255).astype(np.uint8)
    cv2.imwrite(f"prediction.jpg", predicted_image)

Here’s a test image (there’s a car in the center):

Here’s my prediction for the test image:

As you can see, my prediction is just a monochrome mess, not a car shape in the center. Are there any mistakes in my prediction code?

[JordanMakesMaps]

@EtagiBI, nope, nor should it be used with the validation (just training). For validation and testing, just the pre-processing should be applied.

There is such as a thing called Test-Time Augmentation (TTA) which you might want to look into later down the road. It helps to provide slightly higher testing scores.

[JordanMakesMaps]

It doesn’t look like you included the preprocess_input function in the prediction phase. The images you feed to the model during the prediction phase need to follow the exact same format that the images underwent during the training phase.