Problem in accessing my GPU

1,065 views
Skip to first unread message

davi...@tiscali.it

unread,
Nov 15, 2018, 5:20:01 AM11/15/18
to COLMAP
Hi,
I have spotted a possible problem which prevents me to using GPU at all.
My config:
Win10 64bit, Xeon E5-1620 32GB RAM
AMD FirePRo W8100
COLMAP 3.6 Commit 83700eb on 2018-10-31 without CUDA


So this is what happens: Feature extracted always 0 whatever setting, camera model I use.
Processed file [4/15]

 

 
Name: 20181114_183254.jpg

 
Dimensions: 2560 x 1920

 
Camera: #1 - SIMPLE_PINHOLE

 
Focal Length: 2092.00px (Prior)
Features: 0

.. for all images
 I have solved this by disabling GPU: same settings gives me thousands of features extracted.
Then I go to Feature Matching: 

ERROR: Not enough GPU memory to match 2979 features. Reduce the maximum number of matches.

This is actually a misleading message since I have solved by just disabling GPU

Is there any solution to this?

Thanks

oco...@gmail.com

unread,
Feb 4, 2019, 2:12:43 PM2/4/19
to COLMAP
Hi, I have the same issue but whit a diferent hardware>

CPU: Ryzen 5 2600
GPU: Radeon RX 570
Windows 10
16 G of RAM

Using GPU, Colmap does not show errors for on "Feature Extraction" but it find 0 features for all the pictures. If I run it with CPU, it does find arround 10,000 features for picture but it is really slow.

I have also tried it on Linux (Ubuntu) whit the same hardware but even it is able to run, does not use the GPU at all (I use readontop monitor to confirm that the GPU is not used at all).

Also, with feature mathching have exactly the same error:
ERROR: Not enough GPU memory to match 15034 features. Reduce the maximum number of matches.
ERROR: SiftGPU not fully supported

Any idea about how to troublesoot this issue?

Do you thing that compiling the code on this computer could help?

vk2...@gmail.com

unread,
Feb 9, 2019, 8:03:43 PM2/9/19
to COLMAP
COLMAP uses the compute capability of a GPU by using CUDA software .... from NVIDIA .... which software will only run on NVIDIA Graphics cards ...

I would recommend installing a cheap GT1030-based card for testing - available for around $100 - this has CUDA compute capability 6.0 , and installing version 9 of CUDA - note on Windows 10 Visual Studio  is a prerequisite to installing CUDA - for further details see

oco...@gmail.com

unread,
Feb 13, 2019, 12:10:27 AM2/13/19
to COLMAP
Hi, first of all, thanks for your response.

I understand that Colmap just needs CUDA for dense reconstruction but not for feature extraction or feature matching.

In fact, I have successfully tested Colmapt for extraction, matching, and mapping (sparse reconstruction) and use openmvs for dense reconstruction and meshing on a macbook without a graphics card, it has Intel processor that include "intel iris" that is an integrated graphics but not CUDA. 

Regards.


El jueves, 15 de noviembre de 2018, 4:20:01 (UTC-6), davi...@tiscali.it escribió:
Reply all
Reply to author
Forward
0 new messages