Poor predictions from pre-trained model?

[MeyerDLevy]

Hi there,

I suspect this is my fault -- I'm having trouble getting good predictions from either the NYU or KITTI pre-trained models. I load the models the same way they are loaded in test.py:

model2 = keras.models.load_model("C:\\Users\\meyer\\Dropbox\\nga\\python\\github\\DenseDepth\\nyu amazon.h5", custom_objects = {"BilinearUpSampling2D": BilinearUpSampling2D}, compile = False)

and then make my predictions this way, where rgbA is a 480x640 RGB image represented as an array:

z = model2.predict(rgbA)

but my predictions are clearly wrong, as I show in the attached image.

So... does anyone have any idea how I can get accurate predictions out of the network?

Thanks!

[QijinXu]

the pre-trained model is missing, could you please share it?