CODEWITHSUNDEEP
matrixtensorflow
CAFEBABE
CAFEBABE

Reputation: 4101

3D rotation matrix in tensor flow

i would like to learn the parameters of a rotation matrix in 3D using Tensorflow. Therefore, I defined the rotation matrix the following way

g = tf.Graph()
with g.as_default():
  #rotations

  thetax =  tf.Variable(tf.zeros([1]))
  thetax =  tf.Variable(tf.zeros([1]))
  thetay =  tf.Variable(tf.zeros([1]))
  p =  tf.placeholder(tf.float32, [3])
  rotation_matrix_x = tf.pack([tf.constant(1.0),tf.constant(0.0),tf.constant(0.0),
                               tf.constant(0.0),tf.cos(thetax), -tf.sin(thetax),
                               tf.constant(0.0),tf.sin(thetax), tf.cos(thetax)])
  rotation_matrix_y = tf.pack([
                          tf.cos(thetax),tf.constant(0.0), -tf.sin(thetax),
                          tf.constant(1.0),tf.constant(0.0),tf.constant(0.0),
                          tf.sin(thetax),0, tf.cos(thetax)])


 rotation_matrix_z = tf.pack([
                              tf.cos(thetax), -tf.sin(thetax),tf.constant(0.0),  
                              tf.sin(thetax), tf.cos(thetax),tf.constant(0.0),
                              tf.constant(1.0),tf.constant(0.0),tf.constant(0.0)])
 rotation_matrix_x = tf.reshape(rotation_matrix_x, (3,3))
 rotation_matrix_y = tf.reshape(rotation_matrix_y, (3,3))
 rotation_matrix_z = tf.reshape(rotation_matrix_z, (3,3))
 rotated = tf.mult(tf.mult(rotation_matrix_x,tf.mult(rotation_matrix_y,rotation_matrix_z) ,p)

I have now two problems

  1. I get an error message: ValueError: Shapes TensorShape([]) and TensorShape([Dimension(1)]) must have the same rank
  2. is there a more elegant way to define the rotation matrix which does not introduce any extra degree of freedom? E.g., Normal vector + angle would be perfectly acceptable.

Thanks in advance

Upvotes: 1

Views: 1675

Answers (2)

ndrplz
ndrplz

Reputation: 1644

I recently run across the same problem. This is my current solution:

one  = tf.ones_like(cos_rot_x,  dtype=tf.float32)    
zero = tf.zeros_like(cos_rot_x, dtype=tf.float32)

rot_x = tf.stack([tf.concat([one,        zero,      zero],      axis=1),
                  tf.concat([zero,       cos_rot_x, sin_rot_x], axis=1),
                  tf.concat([zero,      -sin_rot_x, cos_rot_x], axis=1)], axis=1)

rot_y = tf.stack([tf.concat([cos_rot_y,  zero,     -sin_rot_y], axis=1),
                  tf.concat([zero,       one,       zero],      axis=1),
                  tf.concat([sin_rot_y,  zero,      cos_rot_y], axis=1)], axis=1)

rot_z = tf.stack([tf.concat([cos_rot_z,  sin_rot_z, zero],      axis=1),
                  tf.concat([-sin_rot_z, cos_rot_z, zero],      axis=1),
                  tf.concat([zero,       zero,      one],       axis=1)], axis=1)

rot_matrix = tf.matmul(rot_z, tf.matmul(rot_y, rot_x))

Notice that in this snippet cos_rot_x has shape (batchsize, 1) so you can keep the batch dimension during the transformation.

Upvotes: 3

mrry
mrry

Reputation: 126154

For problem (1)—the shape error—I think the problem is caused by the fact that you are trying to pack together scalars (such as tf.constant(0.0)) with single-element vectors (i.e. tf.Variable(tf.zeros([1]))). You should be able to fix this by redefining the variables as scalars:

thetax = tf.Variable(tf.zeros([]))
thetax = tf.Variable(tf.zeros([]))
thetay = tf.Variable(tf.zeros([]))

I'm not sure about how to redefine the problem more elegantly... but hopefully this gets you unstuck!

Upvotes: 2

Related Questions