[4eyes] LSD-SLAM

[Tobias Hollerer]

Forgot to mention that Schoeps' paper won Best Short Paper at ISMAR.

I'm right now at an overview talk by Jakob Engel. Yes, it is in 
principle based on methods from the 1990s, but there are many different 
interesting details to attend to to make it work well. I would say the 
faster hardware made the re-visitation of direct alignment methods 
attractive again, but there are definitely nice additional insights 
here. Have you looked at the sim(3) based pose-graph estimation they're 
doing for direct tracking with scale? Looked to me like you had an edge 
there...

Cheers,

     Tobias


On 9/15/14, 12:32 AM, Tobias Hollerer wrote:
>
> Yes, I attended Thomas Schoeps' talk. You'll meet him very soon, as he 
> will start a PhD with Marc Pollefeys imminently.
> The system appear to work very well on Smartphones (the focus of the 
> ISMAR paper, attached):
>
> https://www.youtube.com/watch?src_vid=GnuQzP3gty4&v=X0hx2vxxTMg
>
> I also thought that it would be a good alternative basis for Steffen's 
> work.
>
> Thomas says that he is not disinclined to also making the mobile code 
> base available, but he won't have time to document and maintain it.
>
> Cheers,
>
>     Tobias
>
> On 9/14/14, 7:04 PM, Chris Sweeney wrote:
>> Here are a bunch of videos of LSD-SLAM: 
>> https://www.youtube.com/watch?v=isHXcv_AeFg&list=UURf1mhfcDeJS2HLP6YR07kA 
>> Like I said, the results are undeniably awesome.
>>
>> It's clear that this semi-dens modeling would be great for something 
>> like Steffen's work because you can use the depth estimates to form a 
>> collision mesh over the scene (similar to the triangulated mesh I 
>> believe Steffen uses).
>>
>> They also had a paper at ISMAR on the same system. I would be curious 
>> to hear impressions from anybody who attended.
>>
>> On Thu, Sep 11, 2014 at 10:52 AM, Chris Sweeney 
>> <cmsweeney at cs.ucsb.edu <mailto:cmsweeney at cs.ucsb.edu>> wrote:
>>
>>     To summarize the paper: they perform SLAM using images directly
>>     without any features. They do this by performing an optimization
>>     that aligns images in areas of high gradient intensity. Depth
>>     estimates at these locations are optimized and the uncertainty is
>>     recorded. Then they optimize keyframes to include scale in
>>     addition to rotation and translation, so they're more robust to
>>     drift. Any frames that aren't keyframes are used to reduce the
>>     variance in depth estimation.
>>
>>     There were largely mixed opinions of this paper at ECCV and many
>>     were not that impressed. I attended the talk and had a long
>>     discussion with Jakob afterwards as well. Overall, the results are
>>     undeniably awesome and I'm excited that it is open source but I
>>     don't think it is revolutionary by any means. I would argue only
>>     has a tiny novelty. This is almost exactly the same work that the
>>     Sarnoff Group (Michal Irani, David Nister, Anandan, etc.) was
>>     doing in the 90's... but now we have faster computers. The only
>>     addition I can see is tracking scale in keyframes which allows you
>>     to easily change environments (i.e. go from indoor to outdoor)
>>     without scale drift.
>>
>>     However, I would say that people should definitely start using
>>     this over PTAM assuming the source code that is released is
>>     reasonable. It performs faster, gives more accurate depth maps,
>>     and does not require multiple threads so there is almost no reason
>>     to not use it...
>>
>>
>>     On Thu, Sep 11, 2014 at 9:52 AM, Tobias Hollerer <holl at cs.ucsb.edu
>>     <mailto:holl at cs.ucsb.edu>> wrote:
>>
>>
>>         People in the AR tracking community I respect and trust speak
>>         highly of (or rather have high expectation for) this work
>>         that is being presented at ECCV:
>>
>>         http://vision.in.tum.de/research/lsdslam
>>         Open source license to be announced imminently.
>>
>>         Chris, would you try to catch the talk at ECCV (today?), and
>>         report your impressions to the lab?
>>         Thanks & Cheers,
>>
>>             Tobias
>>
>>         --         Tobias Hollerer
>>         Professor, Department of Computer Science
>>         University of California, Santa Barbara, CA 93106-5110
>>
>>         holl at cs.ucsb.edu <mailto:holl at cs.ucsb.edu>, Office:
>>         (805)284-9395 <tel:%28805%29284-9395>, Fax: (805)893-8553
>>         <tel:%28805%29893-8553>
>>
>>
>>         _______________________________________________
>>         Ilab-users mailing list
>>         Ilab-users at lists.cs.ucsb.edu 
>> <mailto:Ilab-users at lists.cs.ucsb.edu>
>>         https://lists.cs.ucsb.edu/mailman/listinfo/ilab-users
>>
>>
>>
>

-- 
Tobias Hollerer
Professor, Department of Computer Science
University of California, Santa Barbara, CA 93106-5110

holl at cs.ucsb.edu, Office: (805)284-9395, Fax: (805)893-8553