search

JiawangBian / sc_depth_pl

SC-Depth (V1, V2, and V3) for Unsupervised Monocular Depth Estimation Webpage:https://jiawangbian.github.io/sc_depth_pl/
GNU General Public License v3.0
439 stars 70 forks source link

Problems of results on nyu_scv2 #67

Open Cresynia opened 1 year ago

Cresynia commented 1 year ago

Hello! Thanks for sharing your code! And I have successfully run it. But when I ran the sc_depth v2, I couldn't get the corresponding results in the paper, it has a gap. And I also ran the provided model nyu_scv2, it also can't get the corresponding results. I don't know where I was wrong. Could you give me some help? Thanks! image

JiawangBian commented 1 year ago

This is a reimplemented version, so the performance is slightly different from the original paper. If you need better results, you can use a larger batch size and increase training iterations.

Cresynia commented 1 year ago

Thanks for your answer! Could you tell me your default parameters, like the number of batch size, training iterations, some parameter in the config.py. (photo_weight, geometry_weight, smooth_weight), I find it is different from the paper. I want to reproduce it, really thanks!

JiawangBian commented 1 year ago

The current version is well-tuned, so you just need to increase batch size, based on your GPU memory, to improve the performance.