search

d-vo / collab_orb_slam2

Collaborative visual SLAM pipeline as used in Collaborative Visual SLAM using Compressed Feature Exchange (RAL 2018).
https://d-vo.github.io/RAL18/
Other
52 stars 22 forks source link

Is Real Scale recovered with multiple Mono Cam ? #3

Open zeryabmoussaoui opened 5 years ago

zeryabmoussaoui commented 5 years ago

Thanks for sharing your code !

I've (quickly) read the article and the code and i didn't find a mention for recovering the Real Scale. I'm wondering if it is possible in case of multiple monocular cameras.

Many thanks in advance.

soulslicer commented 4 years ago

Does this code work with multiple monocular cameras?

soulslicer commented 4 years ago

@zeryabmoussaoui How did you get it to work with Monocular Datasets from KITTI?

zeryabmoussaoui commented 4 years ago

You just have to run the example provided with collab_orb_slam2

soulslicer commented 4 years ago

Can I please get more information?

According to the docs, it just says that you have to run either the Depth Coding or Stereo Coding example. But we have only a Mono Camera. So how is this possible?

Also, it seems that I will need to compile them with ROS? As they appear to be ROS modules

zeryabmoussaoui commented 4 years ago

I re-checked with the dev team, you were right : we adapted the examples so agents could be monocular cams. The results were'nt good enough and we finally chose another solution for colab SLAM. I hope this will help you !

zeryabmoussaoui commented 4 years ago

This may help you : https://github.com/anitahsuho/collab_orb_slam2

soulslicer commented 4 years ago

Thank you. I can see that it has a Mono Example.

Why were the results not good? If so, what other open source solution did you switch to

zeryabmoussaoui commented 4 years ago

After benchmarking, we chose : https://github.com/VIS4ROB-lab/ccm_slam which recovers fine real scale, which collab_orb_slam2 couldn't

soulslicer commented 4 years ago

Okay, are you using this for Drones or for AV scenarios?

Also, I tried to compile your library, but it seems to have hardcoded paths:

CMake Error: The current CMakeCache.txt directory /home/raaj/collab_orb_slam2/build/CMakeCache.txt is different than the directory /home/anita/slam/dep/collab_orb_slam2/build where CMakeCache.txt was created. This may result in binaries being created in the wrong place. If you are not sure, reedit the CMakeCache.txt
CMake Error: The source "/home/raaj/collab_orb_slam2/CMakeLists.txt" does not match the source "/home/anita/slam/dep/collab_orb_slam2/CMakeLists.txt" used to generate cache.  Re-run cmake with a different source directory.
zeryabmoussaoui commented 4 years ago

Not our implementation, we cloned this repo and had to fix some issues, but i cannot find the local repo. Before trying to help , are you familiar with building ORBSLAM2 ?

soulslicer commented 4 years ago

I have no issues with ORBSLAM2 and have used it a fair bit. I've debugged those errors anyway. The issue was that the build folders were pushed to the repo

Unfortunately, there are quite a few other compiler issues such as protobuf errors

soulslicer commented 4 years ago

@zeryabmoussaoui

Okay I eventually got it to compile. But while the Stereo version works fine, the Mono version segfaults. Looking into the Mono code, it seems to have all sorts of wierd application specific hacks? Is this not a general framework that'll say..work with the KITTI dataset?

For example It'll crash with

Handling robot 0
Segmentation fault (core dumped)
zeryabmoussaoui commented 4 years ago

Sorry to hear about your difficulties, we correct many issues in the code ( including the mono code) , and adapt it for real time application, without any complete rework