[ros-users] Reinventing the wheel: training and evaluation of algorithms for objects detection in images

[Radu Bogdan Rusu]

Rodrigo,


On 11/21/2010 02:10 PM, Rodrigo Benenson wrote:
> Yes indeed I have seen Rein and the other modules in the
> object_recognition stack. These seem all to be "online oriented". Rein
> defines some training methods, but they seem oriented towards online
> training.

There's not much difference in ROS between online and offline training. We record data in BAG files or PCD/PNG (in some 
cases) and play them back to the system, so it really always appears as the system is always learning "online".

> ( included it in the "Similar initiatives and project" of my initial
> draft https://wave.google.com/wave/waveref/googlewave.com/w+yH-HOCb6H
> )

Great! I'll have a look at it later.

> What I am aiming for is to have some kind on standard architecture for
> the offline training stage (and evaluation). This part seems not to be
> covered by any ROS package I am aware of.

The fact that we don't have this yet is an amazing opportunity for you and others to help us out! :) We'd love to work 
with as many people as possible on this, and as you probably noticed, we're all barely just scratching the surface, so 
let's join efforts and make sure that we nail together this within the next few months.

> It seems that Caroline Pantofaru is somewhat working on the
> "exchanging datasets and providing easy evaluation" part of the
> project with the "Moving People, Moving Platform Dataset" initiative,
> but very little of it is visible right now.

Yes, Caroline has done an _amazing_ job on the people dataset. She will bring it out from "the shadows" :) and make it 
available to everyone soon. She's been working hard with some of our colleagues in the community to finish it.

>> Together with your help and the help of the community, we might be able to provide a solution for the evaluation of detectors for object>  recognition.
>
> I'm looking for the best way of creating this tool to make most
> useful. ROS integration would clearly be a nice to have, but as
> mentioned, I am worried it may be "too much" for non-robotics users. I
> guess this will depend on the feedback I get.

We're very happy to write code that's not ROS dependent. OpenCV is already out there, and PCL is written in a way that 
will allow us within a few days to export it outside ROS. REIN follows/will follow a similar pattern.

However, personally I don't think people should think about ROS as "robotic" centric. I myself, am very interested in 
perception methods, but I never think about them as being "robotic only", so I use ROS to ease and simplify my work in 
perception. Besides the fact that the ROS acronym stands for/includes "robotics", the ROS core has nothing to do with 
robotics, and can easily be used for any projects in perception/vision. We do provide a lot of packages in the "ROS 
universe" that are robotics centric, but they are not needed most of the time for tasks such as the one that we're 
trying to solve here.

All in all, think about ROS as an amazing set of tools supported by a great community that can simplify our work. For 
example, ROS-bash has helped me so much in terms of productivity, that I find myself going to other systems which do not 
have ROS installed typing "roscd ~" :) It's like a drug!

> Inputs on the questionnaire are welcome ;P
> https://spreadsheets.google.com/viewform?formkey=dFFzaDlLM1liVGNOS2FENnhrc1VWckE6MQ
>
> Thanks Radu for your comments and let us hope we get to a "good
> enough" solution.
> Regards,
> rodrigob.

I propose that we iterate a bit more on the mailing list before moving completely to Google documents, as it keeps 
things a bit more dynamic and we can involve other people in the discussion. Once we get to some consensus, we can 
create wiki pages and other documents that will keep a centralized archive of our discussion results. $0.02 :)

In any case - thanks for bringing this up and glad to have you onboard!

Cheers,
Radu.