0
$\begingroup$

I would like to teach a model the environment of a room. I'm doing so by mapping a camera pose (x, y, z, q0, q1, q2, q3) to its corresponding image; where x, y, z represent location in Cartesian coordinates and qn represent quaternion orientation. I have tried numerous decoder architectures but I get blurry results with little or no details; as can be seen from the images below:

enter image description here enter image description here

I am using Adam optimizer with a learning rate of 0.0001, and my network architecture is as follows:

  • ReLU(fc(7, 2048))
  • ReLU(fc_residual_block(2048, 2048))
  • ReLU(fc_residual_block(2048, 2048))
  • Reshape
  • ReLU(ConvTransposed2D(in=128, out=128, filter_size=3, stride=2))
  • ReLU(ConvTransposed2D(in=128, out=128, filter_size=3, stride=2))
  • ReLU(ConvTransposed2D(in=128, out=128, filter_size=3, stride=2))
  • ReLU(ConvTransposed2D(in=128, out=128, filter_size=3, stride=2))
  • ReLU(ConvTransposed2D(in=128, out=1, filter_size=3, stride=2))

I have tried different learning rates, loss functions(MSE, SSIM) and even batch normalization. Is there something that I'm missing here?

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.