MVSNet_pytorch icon indicating copy to clipboard operation
MVSNet_pytorch copied to clipboard

Depth refinement

Open ZJCRT opened this issue 5 years ago • 6 comments

Hi there,

great work! Is there any reason why you exclude the depth refinement from the evaluation?

Regards

ZJCRT avatar Jul 09 '19 12:07 ZJCRT

I didn't try with depth refinement, but I think you can try to see the difference.

xy-guo avatar Jul 09 '19 18:07 xy-guo

I had to modify some stuff, but i trains now.

urbste avatar Jul 14 '19 08:07 urbste

Hi,during training process ,the output depth size is 1/4 of input image,the depth refinement module neet concat depth map and input image, so how to use depth refinement module? Upsample depth size or downsample image to keep the same size?

wenmingMeng avatar Oct 25 '19 04:10 wenmingMeng

@wenmingMeng i think downsample image to keep the same size

whubaichuan avatar Jan 10 '20 08:01 whubaichuan

Hello, @urbste Have you tried with depth refinement? From the original MVSNet paper, the refinement network only slightly improves the evaluation results. i.e. from 75.58 to 75.69(<1mm f-score)

XYZ-qiyh avatar May 08 '20 10:05 XYZ-qiyh

es the evaluation results. i.e. from 75.58 t

i think refinement module is useless。

wenmingMeng avatar May 08 '20 10:05 wenmingMeng