I am currently working on system of snn for visual place recognition, I am working with Brian 2, my model is using the Diehl and Cook the system implemented in the article of somayah housiani (her system is the fiḥel and Cook with weighted méthode and modular approach and adding a seqslam in the last), my probleme is that I am having great results when there is lightchanging condition but not with changing of direction of rotation (my dataset is a sequence of image of a rebot moving)
The problème is need to fix this so I can implement the model in application of real time
What you have aready tried
I have tried changing the preetraitement of the image by trying sift or hog or orb none of them solve the probleme
If any advice of this topic I would be glad thank
Hi @Hakim, welcome to the forum! I’ve moved your question to the Science, Projects, and Showcases category, since it is so much about how to use Brian, but more a general modelling question. This type of application is not my domain, so I am afraid I won’t be able to help you with it. As a general suggestion: it is very difficult to help you based on what you wrote on your post. You have a better chance that someone replies if you ask a more specific question, show some of your code and/or results, etc.
Alright, i can share the more of the code by email if anyone is intrested this is the pipeline i am doing :
phase 1 : data input
-transforme image to Grey
-resize to 28 *28
-patch normalization ( the for better the result in change of lighting condition
-sift
phase2 :
-apply rate cooding by poisson groupe
-phase 3 :
using a modular approach using the architecture of diehl and cook ( each snn is learning 25 places )
i changed the way the snn reconized places using a weighted methode ( regulazation , normalisation … )
phase 5 :after the modular snn mode give matrice of similarité it goes to a seqslam of 2 or 4 length
phase 5 : the last prediction
the dataset are places of rebot moving in sequence
my probleme that the model is very sensible the variation ( lets say i trained the model on image 1 and 8 of the sequece when try the model in real time application i fails the reconized the place it must be the exact angle of image 1 or 8 if there is a small rotation or not the same angle it fails to give me good resulats ( after adding the sift it gave me better resulats but still not satisfied resulats )