ros-perception dead?
35 views
Skip to first unread message
Joshua Whitley
Nov 15, 2018, 3:59:22 PM11/15/18
to ROS Perception Special Interest Group
This SiG seems like it hasn't seen conversation since August of last year and the GitHub group definitely needs some love (almost 100 pull requests are dead or un-reviewed across multiple repos). Is anyone still out there? Do we need to recruit some new maintainers?
G.A. vd. Hoorn - 3ME
Nov 16, 2018, 3:26:46 AM11/16/18
to ros-sig-p...@googlegroups.com
Gijs
v4hn
Nov 16, 2018, 8:55:29 AM11/16/18
to ros-sig-p...@googlegroups.com
Hi there,
Migrating the list could modernize it somewhat, but does not change the fact
that few people contribute to the ROS-specific perception projects/code bases.
Please someone contradict me on this, but...
A lot of the current code even originates from before the "Big Neural Network Revolution",
take ORK as an example, and some parts could probably be replaced by using
other third-party frameworks and more "modern" approaches.
This is quite a lot of work and the perception research-community who implements
current state-of-the-art algorithms is rather orthogonal to the ROS community
who would be interested in *utilizing* perception modules.
At times it's even hard to talk to each other.
It seems to me at the moment all manipulation startups/research groups implement their own,
lab-owned solutions for perception and while many of them share similar approaches,
the implementations are fully independent - and often corporate IP/closed source.
The most common basis is sometimes OpenCV, in some cases only Eigen because
someone grew tired of OpenCV API changes, its bugs and non-beginner-friendly documentation.
I know a number of groups who shy PCL because it became more and more bloated over time.
On the other hand Tensorflow, Keras, Torch, <name your learning framework>...
see more and more use here.
If you talk to speakers on robotic perception topics at robotics conferences about
what specific code you could use, they often point to their own lab's implementation
which usually has quite a few limitations on the objects/ sensors, robotic setup,
and/or hardware resources.
There is potential to improve this sad state again, looking at things like
https://rse-lab.cs.washington.edu/projects/posecnn/ , but there is no real common system
in the ROS ecosphere yet that most people would "just use".
For the research community it does not get papers published to improve and polish
promising research projects to be usable in general scenarios.
For industry it's just too good corporate IP to publish open source
(and probably it would have to be polished just as well).
v4hn
> --
> You received this message because you are subscribed to the Google Groups "ROS Perception Special Interest Group" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to ros-sig-percept...@googlegroups.com.
> To post to this group, send email to ros-sig-p...@googlegroups.com.
> To view this discussion on the web visit https://groups.google.com/d/msgid/ros-sig-perception/c66f32ab-a6df-0414-af85-2431911b1edb%40tudelft.nl.
> For more options, visit https://groups.google.com/d/optout.
>
--
Michael 'v4hn' Görner, M.Sc. Cognitive Science, PhD Student
Universität Hamburg
Faculty of Mathematics, Informatics and Natural Sciences
Department of Informatics
Group Technical Aspects of Multimodal Systems
Vogt-Kölln-Straße 30
D-22527 Hamburg
Room: F-315
Phone: +49 40 42883-2432
Website: https://tams.informatik.uni-hamburg.de/people/goerner/
Migrating the list could modernize it somewhat, but does not change the fact
that few people contribute to the ROS-specific perception projects/code bases.
Please someone contradict me on this, but...
A lot of the current code even originates from before the "Big Neural Network Revolution",
take ORK as an example, and some parts could probably be replaced by using
other third-party frameworks and more "modern" approaches.
This is quite a lot of work and the perception research-community who implements
current state-of-the-art algorithms is rather orthogonal to the ROS community
who would be interested in *utilizing* perception modules.
At times it's even hard to talk to each other.
It seems to me at the moment all manipulation startups/research groups implement their own,
lab-owned solutions for perception and while many of them share similar approaches,
the implementations are fully independent - and often corporate IP/closed source.
The most common basis is sometimes OpenCV, in some cases only Eigen because
someone grew tired of OpenCV API changes, its bugs and non-beginner-friendly documentation.
I know a number of groups who shy PCL because it became more and more bloated over time.
On the other hand Tensorflow, Keras, Torch, <name your learning framework>...
see more and more use here.
If you talk to speakers on robotic perception topics at robotics conferences about
what specific code you could use, they often point to their own lab's implementation
which usually has quite a few limitations on the objects/ sensors, robotic setup,
and/or hardware resources.
There is potential to improve this sad state again, looking at things like
https://rse-lab.cs.washington.edu/projects/posecnn/ , but there is no real common system
in the ROS ecosphere yet that most people would "just use".
For the research community it does not get papers published to improve and polish
promising research projects to be usable in general scenarios.
For industry it's just too good corporate IP to publish open source
(and probably it would have to be polished just as well).
v4hn
> You received this message because you are subscribed to the Google Groups "ROS Perception Special Interest Group" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to ros-sig-percept...@googlegroups.com.
> To post to this group, send email to ros-sig-p...@googlegroups.com.
> To view this discussion on the web visit https://groups.google.com/d/msgid/ros-sig-perception/c66f32ab-a6df-0414-af85-2431911b1edb%40tudelft.nl.
> For more options, visit https://groups.google.com/d/optout.
>
--
Michael 'v4hn' Görner, M.Sc. Cognitive Science, PhD Student
Universität Hamburg
Faculty of Mathematics, Informatics and Natural Sciences
Department of Informatics
Group Technical Aspects of Multimodal Systems
Vogt-Kölln-Straße 30
D-22527 Hamburg
Room: F-315
Phone: +49 40 42883-2432
Website: https://tams.informatik.uni-hamburg.de/people/goerner/
Joshua Whitley
Dec 7, 2018, 2:12:05 PM12/7/18
to ROS Perception Special Interest Group
Looks like this Google group was deprecated a while ago in favor of the Discourse group here: https://discourse.ros.org/c/perception
I'll post something similar there as well. Thanks!
> To unsubscribe from this group and stop receiving emails from it, send an email to ros-sig-perception+unsub...@googlegroups.com.
Reply all
Reply to author
Forward
0 new messages