Created
February 23, 2017 09:49
-
-
Save geohot/1f6475bbd3c75f3621acf2c6ea04fc0b to your computer and use it in GitHub Desktop.
Really Low Quality Bundle Adjustment in Tensorflow
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
guess = np.concatenate([np.array(map(lambda x: np.array(x[:,3]).ravel(), rts)), fpts[:, 0:3]], axis=0) | |
selects = [] | |
xys = [] | |
for i, trk in enumerate(good_good_tracks): | |
for shot, kp in trk: | |
xy = G.node[(shot,kp)]['kp']/0.81412059 | |
xys.append(xy) | |
select = np.zeros((len(frms)+len(good_good_tracks),)) | |
select[shot] = 1 | |
select[len(frms)+i] = 1 | |
selects.append(select) | |
selects, xys = np.array(selects), np.array(xys) | |
# initial loss | |
vt = np.dot(selects, guess) | |
proj = (vt[:, 0:2].T / vt[:, 2]).T | |
errs = np.sum((xys - proj)**2, axis=1) | |
print np.sum(errs) | |
# setup tensorflow | |
import tensorflow as tf | |
#fake_guess = np.zeros_like(guess) | |
#fake_guess[len(frms):, 2] = 1.0 | |
t = tf.Variable(guess, dtype=tf.float32) | |
selects_tf = tf.placeholder(tf.float32, [None, len(frms)+len(good_good_tracks)]) | |
xys_tf = tf.placeholder(tf.float32, [None, 2]) | |
mask_tf = tf.placeholder(tf.float32, [None]) | |
vt = tf.matmul(selects_tf, t) | |
proj = tf.transpose(tf.transpose(vt[:, 0:2]) / vt[:, 2]) | |
cost = tf.reduce_sum(tf.transpose(((xys_tf - proj)**2)) * mask_tf) | |
optimizer = tf.train.AdamOptimizer().minimize(cost) | |
init = tf.global_variables_initializer() | |
sess = tf.Session() | |
sess.run(init) | |
mask = np.array([1.0] * (selects.shape[0])) | |
# run for 100 epochs | |
for i in range(100): | |
ret = sess.run([optimizer, cost], feed_dict={selects_tf: selects, xys_tf: xys, mask_tf: mask}) | |
if i%10 == 0: | |
# do ransac | |
vt = np.dot(selects, t.value().eval(session=sess)) | |
proj = (vt[:, 0:2].T / vt[:, 2]).T | |
errs = np.sum((xys - proj)**2, axis=1) | |
mask[np.where(errs>0.00001)] = 0.0 | |
print i, ret, sum(mask) | |
tfout = t.value().eval(session=sess) | |
tfrts = np.zeros((len(frms),3,4)) | |
tfrts[:, :, 3] = tfout[0:len(frms)] | |
tfrts[:, 0, 0] = 1.0 | |
tfrts[:, 1, 1] = 1.0 | |
tfrts[:, 2, 2] = 1.0 | |
tfpts = tfout[len(frms):] | |
tfrts.shape, tfpts.shape | |
#plot(tfrts[:, :, 3]) |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Hi George,
Was googling some words like "tensorflow and bundle adjustment" and find only this page. Just wondering how this "low quality bundle adjustment "is working in your experiments?
Thanks, Bo