1
$\begingroup$

Question on transfer learning object classification (MobileNet_v2 with 75% number of parameters) with my own synthetic data:

I made my own dataset of three shapes: triangles, rectangles and spheres. each category has 460 samples with diferent sizes, dimensions, different wobbles at edges. They look like this:

enter image description here enter image description here enter image description here

I want the network to classify these primitive shapes in other environments as well with different lighting/color conditions and image statistics.

Even though I'm adding random crops, scaling, and brightnesses, at training step 10 it's already at 100% training and validation accuracy. Cross entropy keeps going down though. I'm using tensorflow hub. The performance of the network in the end could be better within other environments (virtual 3d space with such shapes). Also trained and tested for ~ 50 steps to see if the network is overfitting, but that doesn't work too well.

What alterations would you recommend to generalize better? Or shouldn't I train on synthetic data at all to learn primitive shapes? If so, any dataset recommendations?

Thanks in advance

$\endgroup$
2
  • $\begingroup$ im assuming you dont have labels on the actual data your performing inference on? Also in your case you need to make sure the augmentations cover the difference in domains $\endgroup$
    – mshlis
    Nov 22, 2019 at 15:58
  • $\begingroup$ I don't necessarily have the labels there no. Any suggestions on how to make the augmentations cover the difference in domains? Thanks! $\endgroup$ Nov 22, 2019 at 16:31

0

You must log in to answer this question.