def _build_model(self):
"""Build our MLP network."""
with tf.variable_scope("Matchnet", reuse=tf.AUTO_REUSE):
# For determining the runtime shape
x_shp = tf.shape(self.x_in)
# -------------------- Network archintecture --------------------
# Import correct build_graph function
from archs.cvpr2018 import build_graph
# Build graph
print("Building Graph")
self.logits = build_graph(self.x_in, self.is_training, self.config)
# ---------------------------------------------------------------
# Turn into weights for each sample
self.w = tf.nn.relu(tf.tanh(self.logits)) # bs, n ,2
def _build_model(self):
"""Build our MLP network."""
with tf.variable_scope("Matchnet", reuse=tf.AUTO_REUSE):
# For determining the runtime shape
x_shp = tf.shape(self.x_in)
# -------------------- Network archintecture --------------------
# Import correct build_graph function
arch = self.config.net_arch
if arch == "cvpr_2018":
import archs.cvpr2018 as arch
elif arch == "nips_2018_nl":
import archs.nips2018_nl as arch
# Build graph
print("Building Graph")
self.logits = arch.build_graph(self.x_in, self.is_training, self.config)
# ---------------------------------------------------------------
# Turn into weights for each sample
weights = tf.nn.relu(tf.tanh(self.logits))
# Make input data (num_img_pair x num_corr x 4)
xx = tf.transpose(tf.reshape(
self.x_in, (x_shp[0], x_shp[2], 4)), (0, 2, 1))
# Create the matrix to be used for the eight-point algorithm
X = tf.transpose(tf.stack([
xx[:, 2] * xx[:, 0], xx[:, 2] * xx[:, 1], xx[:, 2],
xx[:, 3] * xx[:, 0], xx[:, 3] * xx[:, 1], xx[:, 3],
xx[:, 0], xx[:, 1], tf.ones_like(xx[:, 0])
], axis=1), (0, 2, 1))
print("X shape = {}".format(X.shape))
wX = tf.reshape(weights, (x_shp[0], x_shp[2], 1)) * X
print("wX shape = {}".format(wX.shape))
XwX = tf.matmul(tf.transpose(X, (0, 2, 1)), wX)
print("XwX shape = {}".format(XwX.shape))
# Recover essential matrix from self-adjoing eigen
e, v = tf.self_adjoint_eig(XwX)
self.e_hat = tf.reshape(v[:, :, 0], (x_shp[0], 9))
# Make unit norm just in case
self.e_hat /= tf.norm(self.e_hat, axis=1, keep_dims=True)
def _build_model(self):
"""Build our MLP network."""
with tf.variable_scope("Matchnet", reuse=tf.AUTO_REUSE):
# For determining the runtime shape
x_shp = tf.shape(self.x_in)
# -------------------- Network archintecture --------------------
# Import correct build_graph function
from archs.cvpr2018 import build_graph
# Build graph
print("Building Graph")
self.logits = build_graph(self.x_in, self.is_training, self.config)
# ---------------------------------------------------------------
# Turn into weights for each sample
weights = tf.nn.relu(tf.tanh(self.logits))
# Move to the Richard Hartley's normalized system
# Normalization matrix for modifying GT E
x_mean = tf.reduce_mean(self.x_in, axis=2, keepdims=True)
x_norm = self.x_in - x_mean
x_mss = tf.square(x_norm)
root2 = tf.sqrt(tf.constant(2., dtype=tf.float32))
x_rms1 = root2 / tf.reduce_mean(tf.sqrt(
tf.reduce_sum(x_mss[:, :, :, :2], axis=-1, keepdims=True)),
axis=2,
keepdims=True)
x_rms2 = root2 / tf.reduce_mean(tf.sqrt(
tf.reduce_sum(x_mss[:, :, :, 2:], axis=-1, keepdims=True)),
axis=2,
keepdims=True)
x_norm *= tf.concat([x_rms1, x_rms1, x_rms2, x_rms2], axis=-1)
scale1 = x_rms1[:, 0, 0, 0]
dx1 = -x_mean[:, 0, 0, 0] * scale1
dy1 = -x_mean[:, 0, 0, 1] * scale1
scale2 = x_rms2[:, 0, 0, 0]
dx2 = -x_mean[:, 0, 0, 2] * scale2
dy2 = -x_mean[:, 0, 0, 3] * scale2
zeros = tf.zeros_like(scale1)
ones = tf.ones_like(scale1)
self.p1 = tf.reshape(
tf.stack([
scale1, zeros, dx1, zeros, scale1, dy1, zeros, zeros, ones
],
axis=1), (x_shp[0], 3, 3))
self.p2 = tf.reshape(
tf.stack([
scale2, zeros, dx2, zeros, scale2, dy2, zeros, zeros, ones
],
axis=1), (x_shp[0], 3, 3))
xx = tf.transpose(tf.reshape(x_norm, (x_shp[0], x_shp[2], 4)),
(0, 2, 1))
# # Make input data (num_img_pair x num_corr x 4)
# xx = tf.transpose(tf.reshape(
# self.x_in, (x_shp[0], x_shp[2], 4)), (0, 2, 1))
# Create the matrix to be used for the eight-point algorithm
X = tf.transpose(
tf.stack([
xx[:, 2] * xx[:, 0], xx[:, 2] * xx[:, 1], xx[:, 2],
xx[:, 3] * xx[:, 0], xx[:, 3] * xx[:, 1], xx[:, 3],
xx[:, 0], xx[:, 1],
tf.ones_like(xx[:, 0])
],
axis=1), (0, 2, 1))
print("X shape = {}".format(X.shape))
wX = tf.reshape(weights, (x_shp[0], x_shp[2], 1)) * X
print("wX shape = {}".format(wX.shape))
self.XwX = tf.matmul(tf.transpose(X, (0, 2, 1)), wX)
print("XwX shape = {}".format(self.XwX.shape))