def loss_sirmse_DepthNorm(truth,
pred,
mindepth=args.mindepth,
maxdepth=args.maxdepth):
truth = DepthNorm(truth, maxdepth)
one_shape = tf.ones_like(truth, dtype=tf.float32)
zero_shape = tf.zeros_like(truth, dtype=tf.float32)
thresh = tf.constant(mindepth, dtype=tf.float32)
mask = tf.where(tf.greater(truth, thresh), one_shape, zero_shape)
num_pixels = K.cast(K.sum(mask), tf.float32)
# pred = pred * 1000
pred = DepthNorm(pred, maxdepth)
pred = tf.clip_by_value(pred, mindepth, maxdepth)
log_diff = tf.math.multiply((tf.math.log(pred) - tf.math.log(truth)), mask)
loss_si_rmse = K.sqrt(
K.sum(K.square(log_diff)) / num_pixels -
K.square(K.sum(log_diff)) / K.square(num_pixels))
return loss_si_rmse
def loss_rmse_DepthNorm(truth,
pred,
mindepth=args.mindepth,
maxdepth=args.maxdepth):
truth = DepthNorm(truth, maxdepth)
one_shape = tf.ones_like(truth, dtype=tf.float32)
zero_shape = tf.zeros_like(truth, dtype=tf.float32)
thresh = tf.constant(mindepth, dtype=tf.float32)
mask = tf.where(tf.greater(truth, thresh), one_shape, zero_shape)
num_pixels = K.cast(K.sum(mask), tf.float32)
# pred = pred * 1000
pred = DepthNorm(pred, maxdepth)
pred = tf.clip_by_value(pred, mindepth, maxdepth)
diff = tf.math.multiply(
pred - truth,
mask) / 1000.0 # mapping the distance from millimeters to meters
loss_mse = K.sum(K.square(diff)) / num_pixels
loss_rmse = K.sqrt(loss_mse)
return loss_rmse
def load_label(self, label_path):
label = np.asarray(Image.open(label_path)).astype(np.float32).reshape(
*self.image_shape, 1)
if self.select_label_mode == "baseline1":
label = np.where(label == 0.0, self.max_depth, label)
label = np.clip(label, self.min_depth, self.max_depth)
elif self.select_label_mode == "baseline":
label = np.clip(label, self.min_depth, self.max_depth)
elif self.select_label_mode == "Norm1":
label = np.where(label == 0.0, self.max_depth, label)
label = np.clip(label, self.min_depth, self.max_depth)
label = DepthNorm(label, self.max_depth)
elif self.select_label_mode == "Norm":
label = np.clip(label, self.min_depth, self.max_depth)
label = DepthNorm(label, self.max_depth)
elif self.select_label_mode == "log":
label = np.clip(label, self.min_depth, self.max_depth)
label = np.log(label)
elif self.select_label_mode == "log1":
label = np.where(label == 0.0, self.max_depth, label)
label = np.clip(label, self.min_depth, self.max_depth)
label = np.log(label)
return label