def compute_reproj_err(self):
def l1loss(label, pred, v_weight=None):
diff = tf.abs(label - pred)
#diff = tf.where(tf.is_inf(diff), tf.zeros_like(diff), diff)
#diff = tf.where(tf.is_nan(diff), tf.zeros_like(diff), diff)
div = tf.count_nonzero(diff,dtype=tf.float32)
# div = tf.count_nonzero(diff,dtype=tf.float32)
if v_weight is not None:
diff = tf.multiply(diff, v_weight)
if v_weight is not None:
return tf.reduce_sum(diff)/(tf.count_nonzero(v_weight,dtype=tf.float32)+0.000000001)
else:
return tf.reduce_sum(diff)/(div+0.000000001)
proj_img, wmask, flow = projective_inverse_warp(
self.keyframe_tf,
1.0/tf.squeeze( self.pred_depth, axis=3),
self.pred_pose,
self.intrinsics_tf,
format='eular'
)
self.reproj_loss = l1loss(self.image_tf,
proj_img,
wmask)
def proj_img(self, R, T, image, depth, matK):
filler = tf.constant([0.0, 0.0, 0.0, 1.0], shape=[1, 1, 4])
filler = tf.tile(filler, [1, 1, 1])
transform_mat = tf.concat([R, T], axis=2)
transform_mat = tf.concat([transform_mat, filler], axis=1)
output_img, _, _, _, _ = utlr.projective_inverse_warp(image,
depth,
transform_mat,
matK,
format='matrix')
tf.summary.image('proj' , \
output_img)
tf.summary.image('tgt' , \
image)
self.output_img = output_img