This paper presents results of training a MDE network with an additional canny edge penalty mask on the loss function. This mask adds a constant weight to all pixels within a certain range of the canny edges in the input. This is a great idea for teaching the net to focus on sharp lines and discontinuities, and their examples show results with marked improvement in depth predictions at edges. Luckily for me, they have not uploaded their canny edge detection code. To make training take anything less then 4000 years, I am going to have to write a PyTorch GPU implementation of the canny edge detection algorithm (woohoo). This may or may not be awful - this repo will be a great starting point, but they have not implemented the whole algorithm, such as the hysteresis part.
This paper presents results of training a MDE network with an additional canny edge penalty mask on the loss function. This mask adds a constant weight to all pixels within a certain range of the canny edges in the input. This is a great idea for teaching the net to focus on sharp lines and discontinuities, and their examples show results with marked improvement in depth predictions at edges. Luckily for me, they have not uploaded their canny edge detection code. To make training take anything less then 4000 years, I am going to have to write a PyTorch GPU implementation of the canny edge detection algorithm (woohoo). This may or may not be awful - this repo will be a great starting point, but they have not implemented the whole algorithm, such as the hysteresis part.