Solve detection task without using R-CNN

Hi
I’m struggling with 3d detection problem using only cnn.

I know that the usual detection task should use r-cnn.
However, I have to use only cnn because I do not have the mask (or location) label of the object I want to find, and I have only count of object.

Do you think I can solve the 3d detection problem using only cnn?
(my data is mrimaging)

I am currently using the classification model after modifying for training.
The training loss converges to zero but the validation loss seems not drop. So maybe it is overfitting.:frowning:

Please help me!!
@ptrblck
I need your help…

That’s an intriguing problem … is your dataset publicly available?

No…
It isnt publicly opened.
Do you have any idea??

Yes - It seems that your model is good but is learning the wrong features.

I would try data augmentation on your images - if you’re not already doing that.
It took me a long time to appreciate the power of augmentation.

I don’t know what kind of augmentation will work for you - that requires many experiments.
Luckily, PyTorch has a nice set of functions for that.

I hope it helps!