added binary version of vocabulary to speedup ORB_SLAM2 start

[poine]

I am a noob at both C++ and git, so please be forgiving.

This is a patch to speedup ORB-SLAM2 startup by loading the vocabulary from a binary file instead of ascii.

I tested it on linux on an intel I7 and an arm Exynos5422 (odroid xu4)
ascii | binary
I7 12.06s | 0.2s
arm 77s | 1s

(ORB_SLAM2 was just not usable on the arm with a startup duration of 77s)

I added a line to build.sh to convert Vocabulary/ORBvoc.txt to Vocabulary/ORBvoc.bin

ORB_SLAM2 uses file extension to choose between ascii and binary.

Test with
./Examples/Monocular/mono_tum Vocabulary/ORBvoc.bin Examples/Monocular/TUM1.yaml /data/tum/rgbd_dataset_freiburg1_xyz/
versus
./Examples/Monocular/mono_tum Vocabulary/ORBvoc.txt Examples/Monocular/TUM1.yaml /data/tum/rgbd_dataset_freiburg1_xyz/

Regards

Poine

[hashten]

This works extremely nice straight out of the box! From a loading time of 20.8 sec to 0.48 sec on my 5 year old really slow AMD A6-3410MX. Thanks a bunch, makes testing much more fun!

[hashten]

Works straight away and changed loading time from 20 sec to 0.5 sec!

[poine]

You're very welcome. You make the that time i spend preparing the patch
worth.

On Fri, Sep 16, 2016 at 6:07 PM, Joel Harsten [email protected]
wrote:

@hashten approved this pull request.

Works straight away and changed loading time from 20 sec to 0.5 sec!

—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
#21 (review),
or mute the thread
https://github.com/notifications/unsubscribe-auth/AAGTzzNueT8k7M-DphLGWRcIImzEeADRks5qqr7DgaJpZM4HX6iC
.

[tdnet12434]

Hi @poine ,
Thx u for your patch, this can save a lot of time for doing my thesis!!
I wanna ask a question. How smooth that ORB_SLAM2 running on odroid XU4? I think i will buy one but i not sure about its max frame rate handling.

[poine]

@thanabadee Blsc

It is difficult to answer your question. Framerate depends on the
resolution of your video, the number of extracted features and so on. But
all in all, i don't think you'll get a great framerate with an XU4 (like
maybe a few fps).

If you are concerned about miniaturization and power consumption, i would
advise to take a look at nvidia tegra based boards as you might be able to
take profit of the cuda cores. Parrot has recently released a tegra based
dev board that look very nice:
https://corporate.parrot.com/en/pressrelease/parrots.l.a.m.dunkturnadroneintoasmartrobot

If power consumption is not critical to your application, i would suggest
to make your life easier and use the fastest intel processor you can
afford. This is the path i am following. This was my first robot, with the
XU4: https://goo.gl/photos/WhxoUfWTV6JWFAtn8
This is my current robot: https://goo.gl/photos/17ptWpjAzf2B6gfK7
I am just using an I7 powered laptop on it. I do not regret switching to
the larger robot, which by the way, was not that much more expensive than
the small one (it's an old electric wheelchair). I would call my first
experiment with the small robot premature optimization.

Hth

Poine

On Thu, Oct 6, 2016 at 6:52 PM, Thanabadee Blsc [email protected]
wrote:

Hi @poine https://github.com/poine ,
Thx u for your patch, this can save a lot of time for doing my thesis!!
I wanna ask a question. How smooth that ORB_SLAM2 running on odroid XU4? I
think i will buy one but i not sure about its max frame rate handling.

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
#21 (comment), or mute
the thread
https://github.com/notifications/unsubscribe-auth/AAGTz9s8EKTszJyXd5J30IN-Jes-NHWRks5qxSdCgaJpZM4HX6iC
.

[sunghoon031]

@poine Did you try with localization-only mode on XU4? Or was it with mapping + loop-closing?

[busybyte1]

added binary version of vocabulary to speedup ORB_SLAM2 start (4122702)

For Windows 8.1, Visual Studio 2017 :
Compiler-Error like: Not interpreted as constant

Thirdparty/DBoW2/DBoW2/TemplatedVocabulary.h
Line 1484:
-char buf[size_node];
+char *buf = new char[size_node];

rest works perfect, many thanks ! (0.45sec to load ORBvoc.bin)

[chenzhanjie]

i am a new, i do not know how to convert, the line which your added is what, thank you.

[busybyte1]

i'm not an expert but if you managed to get orb to work you should know what i write about next, otherwise it's useless.

1. open dbow.sln project, which you should have build with cmake, with VisualStudio .
2. in your project-explorer go to external dependencies and open TemplateVocabulary.h.
3. change the lines like in (4122702) and recompile (if not possible you may have to rebuild dbow with cmake and do it again). recompile always both, debug and release version.
4. almost the same do for your orb-slam2 project (not the examples), wich should be build with cmake, for system.cc and bin_vocabulary.cc (search the files in your project-explorer)
5. read the new code to change intensive, theres a function called "save_as_binary". with this function you can save the vocabulary as bin, so you have to implement it somehow in your tum-project.
6. load the voc.txt and save as .bin (watch the path where it's saved, maybe you have to search it i.e. in your tum-project-folder and copy it to the path from the .txt version)
7. change the voc-filename in your mono/rgbd/stereo_tum from .txt to .bin and load it as .bin

hope i could help you

[YumaTheCompanion]

I've checked the file changes with the vanilla version of ORB_SLAM2 on:
Ubuntu 18.04,
CMAKE 3.10.2,
g++7.4.0.

The solution works as expected.

[Fourierw]

After adding the code I completely rebuilt the 'ORB_SLAM2' project , yet the ORBvoc.bin was nowhere to find.
Then I debugged the 'bin_vocabulary' project , it said "Vocabulary loading failure: This is not a correct text file!" , how can I solve this to get the binary file of ORBvoc ?
(when I run the mono_tum project , the ORB.txt file can be read correctly and the entire program was fine)

[jeremysalwen]

Disclaimer: I don't intend to create spam. If the ORB_SLAM2 developers wish me to stop posting, I will cease immediately.

ORB_SLAM2 seems to be inactive, but there are a number of community patches, including yours, which provide important bugfixes or improvements. Therefore I have created a new project ORB_SLAM_COMMUNITY (https://github.com/jeremysalwen/ORB_SLAM_COMMUNITY) to gather together these improvements and continue active development. I have already merged this PR in the ORB_SLAM_COMMUNITY version. We would welcome any other contributions you might have.