Post - Forum - Plus/4 World



	Home Search Games Tapes Covers Cheats Maps Software New Stuff Hall Of Fame HVTC Game Endings Solutions Remakes Publications Magazines Effects Top List Members Groups Features Plus/4 Encyclopedia Hardware Tools Options Forum Home Search Games Tapes Covers Cheats Maps Software New Stuff Hall Of Fame HVTC Game Endings Solutions Remakes Publications Magazines Effects Top List Members Groups Features Plus/4 Encyclopedia Hardware Tools Options Forum	Login

Forum Help

Post Your Message

Previous Messages

Posted By

Mad
on 2024-01-15
10:54:56

Re: Offtopic: Brightness Invariant Feature Tracker

http://www.puls4r.de/downloads/esbift_sourceb.zip Password is "A"

Enough stuff here I hope.. Just tried to preserver the heavily modified sources.. The sources get modified on a massive scale on my computer.
I think it may be the windows setup or something.. I got heavy modifications on a daily even more granular basis.. I hope this is for some time somehow safe..

Thank you for the help!

https://www.youtube.com/watch?v=0PrRU8MssHY

Posted By

Bionic
on 2024-01-04
15:31:33

Re: Offtopic: Brightness Invariant Feature Tracker

I don't know too much about this. But I was wondering whether you tried to submit it as a pull-request to opencv? That way people may actually try it?

Posted By

Mad
on 2023-12-28
16:32:49

Re: Offtopic: Brightness Invariant Feature Tracker

@Murphy Plain and simple. Yes I did compare it to affine KLT and KLT three or 4 years ago (and it was on par so to same as both in terms of precission even subpixel accuracy was almost the same if I remember right.). The algorithm improve further from there. My hardrive got filled with crap for some time so that I almost had only 400 MB left from the I think 900GB or something. I don't develop it currently further. My point is the potential of the algorithm not the algorithm it self. I'm hoping it didn't offend you personally or your peers that talk about it.

My SourceForge Repository got somehow modified I didn't upload the new source yet (it's christmas have to visit my family now), but here some updated videos with a description on what to change that all looks well again. (mainly two "+" turned into a "-")..

https://www.youtube.com/watch?v=U93TMqEN620 shows already the potential to be using ESBIFT in visual AI (visual transformers) as token source.
https://www.youtube.com/watch?v=fr9hu47Gx7s the first "-" already looks a tiny tiny bit better (features "flowed" much more before this change)
https://www.youtube.com/watch?v=zTd8xBHmsPE the second "-" (changed from a "+" again)..
https://www.youtube.com/watch?v=Y_9woyZj6Z0 the 3d patches somehow work again, have to look at it in the evening..

Sorry Murphy and merry XMas (again sorry)

edit: it's working well again: https://www.youtube.com/watch?v=QElZh73zrA8 Happy XMas.. (my clipboard was changed or not updated before.)

New sources to be found here: https://puls4r.de/downloads/esbift_v_with_gesbift.cpp (not sure if all is working as expected.(source is not modified))

new videos:
https://www.youtube.com/watch?v=xGnezK-bZuw GESBIFT
https://www.youtube.com/watch?v=ufEi0BbByZ0 FESBIFT
https://www.youtube.com/watch?v=hAzRJrZq6uM XESBIFT

Posted By

Murphy
on 2023-12-10
11:27:11

Re: Offtopic: Brightness Invariant Feature Tracker

I don't know much about the subject, but the development is interesting. Have you tried to add to an open source end product to test your algorithm against the original?

If you can compare performance and quality, you could show the power of your algorithm in a much more impressive way.

Posted By

Mad
on 2023-12-03
20:00:43

Offtopic: Brightness Invariant Feature Tracker

Hi guys,

I know that's nothing to do here.. But I recently invented and developed a feature tracking algorithm (around 100% by me) which outperforms most of the industrial standard ones, even the best ones so to say. I think it's possible to implement it on a C16/C64 and alike but it may not be 100% realtime then. The main code is about 10 lines of code and even in asm it shouldn't be so much of lines even on the real thing. I am totally failing to promote that algorithm, I just have no clue what to do anymore. The science world is completly closed and locked up (webspace wise), besides AI algorithms doing a better task on that of course already.

You can find a 5000+ lines implementation of the complete algorithm here.. (It's really just around 10 lines the rest is supporting code).

https://sourceforge.net/projects/esbift/ (0 downloads since months)

Some videos here, but quality is very low.. And maybe a non schooled eye will see only noise (like most academics)..

https://www.youtube.com/watch?v=viOji98H3Jo

https://www.youtube.com/watch?v=XtIYvVe_Y6g

https://www.youtube.com/watch?v=RprFFhw7uQM

very simple description in an "AI-based web text editor": simple docu (with some ?token? in the URL)

I will do some propper videos in the future.. And besides this not belonging here, there is the fact that this algorithm really kicks out all other hard math based old ones if it will be developed further.. I just stopped developing it some months ago.. (Due to my desktop PC being under heavy attacks by malware and viruses and so on..)

Have fun.. (And yes, I am that C16 guy and not a robot promoting nonsense..)

If anyone knows where to promote algorithms, related to robotics,.. Doug (Turner) would be happy to see this I suppose..

edit: Please note that tracking is only one part of the current object recognition libraries and research, the jumpings of the feature points are completely normal for any of all the other algorithms in such settings. And this being not C16, yet.. (just for the people disapproving: And that being the only Feature Tracker absolutely approved by ultra very higher as high realms and even even higher realms maybe even even even higher realms than and so on... But not approved by you.)

It's atleast on par with this one: https://en.wikipedia.org/wiki/Kanade%E2%80%93Lucas%E2%80%93Tomasi_feature_tracker whilst having much more advancements than that one.. I started developing this algorithm in a company with people advocating and teaching about that (KLT).. KLT was much slower performance wise (the GPU implementations I looked at) and much less robust (e.g. not brightness invariant, not good at re-locating, etc..) for instance. Metrics where almost the same with the affine KLT when testing it around 3-4 years ago in a draft version.

"Since the early works of Lucas and Kanade [8] and Shi and Tomasi [10], the Kanade-Lucas-Tomasi (KLT) feature tracker has been used as a de facto standard in handling point features in a sequence of images. From the original formulation a wide variety of extensions has been proposed for better performance."