cleinc / bts

From Big to Small: Multi-Scale Local Planar Guidance for Monocular Depth Estimation
GNU General Public License v3.0
635 stars 179 forks source link

Incorrect NYU cropping caused by Indexing issue #61

Closed jungerm2 closed 3 years ago

jungerm2 commented 4 years ago

In get_projection_mask.m from the original nyuv2 toolkit, the mask is defined by:

mask = false(480, 640);
mask(45:471, 41:601) = 1;

Since Matlab indices start at 1 and are inclusive, the numpy equivalent should be [44:471, 40:601] instead. This is used here: https://github.com/cogaplex-bts/bts/blob/e63ccc00a1d44cc417c745714e3cabe58559f054/utils/eval_with_pngs.py#L205

cogaplex-bts commented 4 years ago

@jungerm2 Thanks for correcting us! We will update our code as well as the results.

jungerm2 commented 4 years ago

I was also wondering why you didn't provide any accuracy metrics with smaller thresholds, I know that the 1.25^i is the standard one, but the results are so good you could've benefited from smaller deltas. Did you consider this? Thanks!