0
$\begingroup$

I'm training a CNN network to find coordinates of an object place on a 2d grid space. The problem is that the error is not going below 3 cm tolerance. My data set consists of 1000 images (240x135 resolution). I'm using a model with 7 conv-maxpool layers and then 5 dense layers (all relu activations).

How can I reduce error further ??

I have tried training for longer durations, used mini-batch GD, used l2 regularization, dropout..

Sample Image

$\endgroup$
2
  • $\begingroup$ It sounds like you have WAY too many parameters for only 1000 images. What is your parameter count? $\endgroup$ – Dave Feb 19 '20 at 19:40
  • $\begingroup$ Trainable parameters = 34237536, considering all weights of convnets and dense layers. $\endgroup$ – Ariyan G Feb 20 '20 at 16:41