I am new to deep learning and fastai. Recently i had been training Unet architecture (with pre-trained resnet50) on the KITTI Dataset for monocular depth estimation.
The input to the network are rgb images, whereas, the groundtruth / label images are the lidar scans. The ground images are gray scale images having 3 identical channels (i.e. same value of a pixel in all the 3 channels).
The task is to let the network learn to predict depth estimates when it is given a rgb image as input.
After training the model, the results obtained on validation set using ‘model.show_results()’ are pretty much as i desire, i.e. a gray scale image which has 3 identical channels (i.e. same value of a pixel in all the 3 channels). However, when i use this trained model for inference, the images obtained are colored images with different values in all the three channels.
I am confused. As per my little knowledge, ideally the prediction from the trained model should have been the same as results obtained after the training.
I will highly appreciate if someone can guide me through this. And How can i get the same results using ‘learn.predict’ as previously obtained from ‘model.show_results()’.
Thanking in advance!