search

introlab / rtabmap

RTAB-Map library and standalone application
https://introlab.github.io/rtabmap
Other
2.8k stars 787 forks source link

Compiling rtabmap with libtorch 1.7.0 #633

Closed jsseng closed 3 years ago

jsseng commented 3 years ago

I am trying to compile the latest rtabmap with libtorch 1.7.0. I am getting this error regarding C++14:

In file included from /home/jseng/Downloads/libtorch/include/torch/csrc/api/include/torch/torch.h:3:0,
                 from /home/jseng/rtabmap/corelib/src/superpoint_torch/SuperPoint.h:9,
                 from /home/jseng/rtabmap/corelib/src/Features2d.cpp:48:
/home/jseng/Downloads/libtorch/include/torch/csrc/api/include/torch/all.h:4:2: error: #error C++14 or later compatible compiler is required to use PyTorch.
 #error C++14 or later compatible compiler is required to use PyTorch.

I am using gcc 7.5.0 (Ubuntu 18.04) and I thought C++14 was standard. I also tried setting -std=C++14, but I am not sure if it is being reset somewhere.

matlabbe commented 3 years ago

Try with with a newest cmake version (at least 3.16).

jsseng commented 3 years ago

That did the trick! Can I ask a question relating to SuperPoint run time? I am now getting update times of greater than 1 sec. with SuperPoint. I believe it is using the GPU when I use the command to monitor usage: nvidia-smi -l 1 (I can see rtabmap_ros). Does a typical forward pass through SuperPoint take this long typically? Thanks!

matlabbe commented 3 years ago

Which kind of GPU are you using? On my GTX 1650, I get more like 300-350 ms per frame. On CPU, I can see over 1 second though.

jsseng commented 3 years ago

I am using an old K620. I timed the forward pass of the network and it took about 600ms. I am using images from a Zed 2 (1280x720). When I increased the SuperPoint threshold, it brought the overall runtime down (around 800ms total including NMS). I have been trying it on a Jetson Xavier and I was getting similar numbers at max power mode. What resolution images do you use?

jsseng commented 3 years ago

I reduced my resolution and it runs faster. I will go ahead and close this.

matlabbe commented 3 years ago

In this video, loop closure detection was done with SuperPoint features using MyntEyeS images of resolution 752x480.