Hi !
I’m trying to train a Resnet-based model (pretrained on Imagenet) on the NYU dataset for depth estimation. I’m only using the small, annotated part of the NYU dataset (which is 1.5k images) but the results are very blurry. I can’t figure out what’s wrong with it. I’m using a loss function composed of the mse, an image gradient loss and a normal surface loss but so far, not luck. I know 1.5k datapoint is rather small, but I wanted to see whether I could overfit before focusing on generalization.
I’ve made a colab notebook and I was wondering if someone could take a look and help me out ?
https://colab.research.google.com/drive/1o2mG_J1mzAyY2jY7rS5fXEmfrCLkmxKT
Thanks a lot !