running on custom dataset with no multiple views

[andreaceruti]

Hi, i am currently trying to track grape bunches on a vineyard row. I have a dataset for training a maskrcnn model, so my idea was to just crop the grape bunches from my dataset, and feed it to the siamese network. So the problem is that for every bunch I have a single camera view, instead of multiple ones.
Do you think that deep SORT method is not useful in my case or I can apply deep SORT with your implementation even in my case?

[seahawks8]

Give me an example image. Not sure if this is the right framework for you.

[andreaceruti]

This is a sample image from the vineyard row I would like to track (so this is frame of a video never seen by my neural network)

and this is a sample crop from my training set.

So to be clear, the dataset which I use for training the network is a set of 200 images representing different vineyards all over the world. Therefore when I have trained my model I pass to it video frames and after the detection part I would like to apply tracking.
What I have done up to now is creating a dataset of 1800 crops that represent different grape bunches.
Instead in this framework, from what I understand, I should build another dataset where I have the same grape bunches seen by different views (in my case they could be different frames). Can i skip this part where I create a new dataset or what? Since it would require a lot amount of time

[vocdex]

@andreaceruti hi there,
I am working on a similar project ( greenhouse plant) and I would like to know if you have succeeded in training the model?
Do you think the images of the same object in different frames could be used to train the siamese network as you mentioned above?
Thanks!