running on custom dataset with no multiple views of the same object

[andreaceruti]

Hi, my problem is that I would like to generate this appearance vector using a dataset that does not contains multiple views of the same object (for example the VeRI dataset), but it contains a single view for every object. So to be more clear I can collect in a dataset 1800 crops and single views of all of them.
Is this the right framework to use or should I choose another tracking framework?

[MMNavetty]

Hello @andreaceruti,
As i understand the algorithm, mutliple view is not necessary to train it. Its purpose is to make a comparison beetween two images and make features describing the images as output. Its goal is to make similar feature when individuals in images are similar.
So with Veri we have individual cars in different images and it completely work !
I hope my explanation are understandable :)
Regards,
MM

[andreaceruti]

@MMNavetty Yes I have understand, thank you.
My problem is that I am trying to capture grape bunches and so features will probably be very similar between them. In my crops what differs a lot is that I have both white and red varieties, but in the video that I am using for tracking the variety is only red.
So I can't tell if applying this framework I will get a result better than SORT, but I will try anyway.

[MMNavetty]

Hello @andreaceruti,
if you have some natural grape bunches that differs in shape, it could work ! try and test it with your dataset and you will see if the algorithm can make the difference. If not, try SORT ! Altough if you have good detection, sort can be enough.
Good luck with your project,
MM

[PHCLenzi]

Hello, @andreaceruti. Can you tell me which version of Tensorflow, Python and other packages you used to run the training? I am having problems running a training with the MARS dataset itself. Did you install these packages with pip or with a conda manager? Thank you very much for your attention!

[MMNavetty]

tensorflow 1.14
python 3.7.11
other packages are billion so i can't annotate everything ! i don't know how to mp in git but i can share a requirement.txt file with you if you really need it