Tag detection is very slow (CUDA 11.1, Windows 10)

Question

Tag detection is very slow (CUDA 11.1, Windows 10)

antithing opened this issue 3 years ago · comments

antithing commented 3 years ago

I have this code built under windows, with CUDA 11.1. The lib compiles, and so does teh 'detection' application.

When i run the detection on the sample image using:

detection -n 3 -i sample/01.png

It finds the markers and displays them, but it takes:

Total time: 5.103234s wall, 5.218750s user + 0.000000s system = 5.218750s CPU (102.3%)

more than five seconds! is this expected? I am looking to run this code in realtime (50fps)

Am i doing something wrong? Can i speed up the tag detection?

Thank you.

Simone Gasparini · Answer 1 · Mon Aug 30 2021 21:07:41 GMT+0800 (China Standard Time)

try detection -n 3 -i sample/01.png --use-cuda
see here
https://github.com/alicevision/CCTag/blob/develop/src/applications/detection/CmdLine.cpp#L44

antithing · Answer 2 · Mon Aug 30 2021 22:19:31 GMT+0800 (China Standard Time)

Ah thank you! This has it running at around 130ms for the sample image, on a GTX 1080.

Aside from resizing the image, is there any other tweaks to make to speed up the functions?

Thanks again!

Simone Gasparini · Answer 3 · Mon Aug 30 2021 22:53:53 GMT+0800 (China Standard Time)

try with a video, because on a single image there might be some overheads involved I guess

antithing · Answer 4 · Mon Aug 30 2021 23:19:50 GMT+0800 (China Standard Time)

Okay thanks, i will try it.

Just so i hav realistic expectations, what is the top speed of this code, roughly, at 1920 x 1080 image size?

Simone Gasparini · Answer 5 · Mon Aug 30 2021 23:22:03 GMT+0800 (China Standard Time)

normally it should be real-time with FullHD videos

antithing · Answer 6 · Wed Sep 01 2021 00:07:21 GMT+0800 (China Standard Time)

Hi again, I am testing with a video file (1920 x 1080) with three markers visible. Using CUDA, on a GTX 1080 card, I still see slow frame times (150 - 230 ms)

The log is:

`#frame 333
Detected 5 candidates
660.546 667.647 0 1
1003.29 682.542 9 1
1354.46 698.786 14 1
1354.4 698.556 14 -4
1354.6 698.46 14 -4

3 markers detected and identified
Times passed in milliseconds: 213.893
Loading image 334 into TagPipe 0
D:\Markers\CCTag\CCTag-develop\src\cctag\Detection.cpp:861
Found more than 60 (62) markers
D:/Markers/CCTag/CCTag-develop/src/./cctag/cuda/pinned_counters.cu:134 called from D:\Markers\CCTag\CCTag-develop\src\cctag\CCTag.cpp:109 Hard-coded number of Nearyby Points in pinned memory is too small.
Increase and recompile.
D:/Markers/CCTag/CCTag-develop/src/./cctag/cuda/pinned_counters.cu:134 called from D:\Markers\CCTag\CCTag-develop\src\cctag\CCTag.cpp:109 Hard-coded number of Nearyby Points in pinned memory is too small.
Increase and recompile.

==== Uploading 62 tags ====
Total time: 0.197040s wall, 0.468750s user + 0.062500s system = 0.531250s CPU (269.6%)`

Is there anything I can do to speed this up? I am aiming for 60fps...

Thank you!

github-actions · Answer 7 · Sat Aug 27 2022 08:05:38 GMT+0800 (China Standard Time)

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.