MiDaS V3.1 (AI monocular depth estimation) VS LeiaPix Converter (Boosting Monocular Depth)

MiDaS V3.1 is out! In pre-release form.

Yes, everybody's favorite monocular depth map generator is here with a new version, better than ever.

A cool thing is that, with this new version, the depth maps come out in the same size as the input images. Was not the case before. A big improvement in terms of ease to use.

I compare this new MiDaS version with the LeiaPix Converter, a (free) website to generate depth maps from single images using the Boosting Monocular Depth machinery under the hood.

Best model, BEiTLarge 512, with resolution 512x512, is on average about 28% more accurate than MiDaS v3.0

Link to MiDaS V3.1 google colab (thanks to Rashad): https://colab.research.google.com/dri...

Link to the depth player: https://depthplayer.ugocapeto.com/