def run():
hyperparams.announceEval()
eval_pairs = hyperparams.getEvalDataGen()
graph, sess = hyperparams.modelFromCheckpoint()
forward_passer = hyperparams.getForwardPasser(graph, sess)
fp_cache = system.ForwardPassCache(forward_passer)
for pair_i in range(len(eval_pairs)):
print('%d/%d' % (pair_i, len(eval_pairs)))
pair = eval_pairs[pair_i]
print(pair.name())
fps = [fp_cache[im] for im in pair.im]
assert len(fps[0]) == FLAGS.num_test_pts
assert len(fps[1]) == FLAGS.num_test_pts
matched_indices = system.match(fps)
inliers = system.getInliers(pair, fps, matched_indices)
matched_fps = [fps[i][matched_indices[i]] for i in [0, 1]]
evaluate.renderMatching(pair, fps, matched_fps, inliers)
def plot(label=None, k=10):
hyperparams.announceEval()
eval_pairs = hyperparams.getEvalDataGen()
# Very special case lfnet:
if FLAGS.baseline == 'lfnet':
x = [50, 100, 150]
y = []
for num_pts in x:
forward_pass_dict = baselines.parseLFNetOuts(eval_pairs, num_pts)
success = np.zeros(len(eval_pairs), dtype=bool)
for pair_i in range(len(eval_pairs)):
pair = eval_pairs[pair_i]
folder, a, b = pair.name().split(' ')
forward_passes = [
forward_pass_dict['%s%s' % (folder, i)] for i in [a, b]
]
matched_indices = system.match(forward_passes)
inliers = system.getInliers(pair, forward_passes,
matched_indices)
if np.count_nonzero(inliers) >= k:
success[pair_i] = True
y.append(np.mean(success.astype(float)))
plt.plot(x, y, 'x', label='%s: N/A' % (hyperparams.methodString()))
else:
pair_outs = cache_forward_pass.loadOrCalculateOuts()
if FLAGS.num_scales > 1 and FLAGS.baseline == '':
fps = [[multiscale.forwardPassFromHicklable(im) for im in pair]
for pair in pair_outs]
else:
fps = [[system.forwardPassFromHicklable(im) for im in pair]
for pair in pair_outs]
pairs_fps = zip(eval_pairs, fps)
stats = [
evaluate.leastNumForKInliers(pair_fps[0], pair_fps[1], k)
for pair_fps in pairs_fps
]
x, y = evaluate.lessThanCurve(stats)
auc = evaluate.auc(x, y, 200)
plt.step(x, y, label='%s: %.2f' % (hyperparams.label(), auc))
return auc
def step(sess):
dp = train_pairs.getRandomDataPoint()
if FLAGS.augment and FLAGS.tds == 'hp':
dp = augment.augmentHpatchPair(dp)
# Feedforward
fps = [forward_passer(im) for im in dp.im]
if type(fps[0]) == multiscale.ForwardPass:
flat = [i.toFlatForwardPass() for i in fps]
else:
flat = fps
matched_indices = system.match(flat)
inliers = system.getInliers(dp, flat, matched_indices)
print('%d inliers' % np.count_nonzero(inliers))
batch, inlier_labels = makeBatchAndInlierLabelsDouble(
dp, fps, matched_indices, inliers)
sess.run(g.train_step,
feed_dict={
g.train_input: batch,
g.inlier_mask: inlier_labels
})
def plot(color=plt.get_cmap('tab10').colors[0]):
n = FLAGS.num_test_pts
hyperparams.announceEval()
eval_pairs = hyperparams.getEvalDataGen()
# Very special case lfnet:
if FLAGS.baseline == 'lfnet':
forward_pass_dict = baselines.parseLFNetOuts(eval_pairs, n)
fps = []
for pair_i in range(len(eval_pairs)):
pair = eval_pairs[pair_i]
folder, a, b = pair.name().split(' ')
forward_passes = [
forward_pass_dict['%s%s' % (folder, i)] for i in [a, b]
]
matched_indices = system.match(forward_passes)
fps.append([forward_passes[i][matched_indices[i]] for i in [0, 1]])
else:
pair_outs = cache_forward_pass.loadOrCalculateOuts()
if FLAGS.num_scales > 1 and FLAGS.baseline == '':
raise NotImplementedError
else:
fps = []
full_fps = []
for pair in pair_outs:
reduced = [
system.forwardPassFromHicklable(i).reducedTo(n)
for i in pair
]
full_fps.append(reduced)
matched_indices = system.match(reduced)
fps.append([reduced[i][matched_indices[i]] for i in [0, 1]])
pairs_fps = zip(eval_pairs, fps)
masks_errs = [
evaluate.p3pMaskAndError(pair_fps[0], pair_fps[1])
for pair_fps in pairs_fps
]
if FLAGS.baseline == '':
for mask_e, pair_fps, full in zip(masks_errs, pairs_fps, full_fps):
mask, rerr, terr = mask_e
pair, fps = pair_fps
evaluate.renderMatching(pair, full, fps, mask)
ninl = np.array([np.count_nonzero(i[0]) for i in masks_errs])
rerrs = np.array([i[1] for i in masks_errs])
rerrs[ninl < 10] = np.inf
terrs = np.array([i[2] for i in masks_errs])
terrs[ninl < 10] = np.inf
if FLAGS.baseline != 'sift':
rlabel, tlabel = hyperparams.label(), None
else:
rlabel, tlabel = '%s, rot.' % hyperparams.label(), \
'%s, transl.' % hyperparams.label()
x, y = evaluate.lessThanCurve(rerrs)
plt.semilogx(x, y, '--', color=color, label=rlabel)
x, y = evaluate.lessThanCurve(terrs)
plt.semilogx(x, y, ':', color=color, label=tlabel)
#
# You should have received a copy of the GNU General Public License
# along with sips2_open. If not, see <http:#www.gnu.org/licenses/>.
import numpy as np
import sips2.flags as flags
import sips2.hyperparams as hyperparams
import sips2.system as system
FLAGS = flags.FLAGS
if __name__ == '__main__':
hyperparams.announceEval()
eval_pairs = hyperparams.getEvalDataGen()
graph, sess = hyperparams.modelFromCheckpoint()
forward_passer = hyperparams.getForwardPasser(graph, sess)
score_wasinlier = []
for pair in eval_pairs:
fps = [forward_passer(im) for im in pair.im]
matched_indices = system.match(fps)
inlier_mask = system.getInliers(pair, fps, matched_indices)
for im_i in [0, 1]:
im = pair.im[im_i]
inlier_indices = matched_indices[im_i][inlier_mask]
scores = fps[im_i].ip_scores
for pt_i in range(len(scores)):
score_wasinlier.append([scores[pt_i], pt_i in inlier_indices])
np.savetxt(hyperparams.wasInlierFilePath(), np.array(score_wasinlier))