# Copyright (C) 2019 Titus Cieslewski, RPG, University of Zurich, Switzerland
# You can contact the author at <titus at ifi dot uzh dot ch>
# Copyright (C) 2019 Konstantinos G. Derpanis,
# Dept. of Computer Science, Ryerson University, Toronto, Canada
# Copyright (C) 2019 Davide Scaramuzza, RPG, University of Zurich, Switzerland
#
# This file is part of sips2_open.
#
# sips2_open is free software: you can redistribute it and/or modify
# it under the terms of the GNU General Public License as published by
# the Free Software Foundation, either version 3 of the License, or
# (at your option) any later version.
#
# sips2_open is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU General Public License for more details.
#
# You should have received a copy of the GNU General Public License
# along with sips2_open. If not, see <http:#www.gnu.org/licenses/>.
from absl import flags
import os
import tensorflow as tf
from rpg_datasets_py.hpatches import HPatches
import rpg_datasets_py.euroc
import rpg_datasets_py.kitti
import rpg_datasets_py.robotcar
import rpg_datasets_py.tum_mono
import baselines
import graph
import multiscale
import sequences
import system
FLAGS = flags.FLAGS
def shortString():
ret = 'd=%d_tds=%s_nms=%d' % (FLAGS.d, FLAGS.tds, FLAGS.nms)
if FLAGS.num_scales > 1:
ret = ret + '_ms%d_%.02f' % (FLAGS.num_scales, FLAGS.scale_factor)
if FLAGS.pbs:
ret = ret + '_pbs'
if FLAGS.augment:
ret = ret + '_aug'
if FLAGS.w != 128:
ret = ret + '_w=%d' % FLAGS.w
if FLAGS.scale_aug_range != 1.:
ret = ret + '_sar=%.01f' % FLAGS.scale_aug_range
if FLAGS.klti:
ret = ret + '_klti'
if FLAGS.lk:
ret = ret + '_lk'
if FLAGS.ol != 0.5:
ret = ret + '_ol=%.2f' % FLAGS.ol
return ret
def announce(what):
print(os.linesep + os.linesep + what + os.linesep + os.linesep)
def announceTraining():
announce('Training %s...' % shortString())
def getTrainDataGen():
if FLAGS.klti:
pair_class = sequences.PairWithIntermediates
else:
pair_class = sequences.PairWithStereo
assert FLAGS.pck == 'tr'
if FLAGS.tds == 'kt':
ks = [rpg_datasets_py.kitti.KittiSeq(i)
for i in rpg_datasets_py.kitti.split('training')]
return sequences.MultiSeqTrackingPairPicker(ks, FLAGS.ol, pair_class)
elif FLAGS.tds == 'rc':
rs = rpg_datasets_py.robotcar.getSplitSequences('training')
return sequences.MultiSeqTrackingPairPicker(rs, FLAGS.ol, pair_class)
elif FLAGS.tds == 'tm':
tms = [rpg_datasets_py.tum_mono.Sequence(i)
for i in ['01', '02', '03', '48', '49', '50']]
return sequences.MultiSeqTrackingPairPicker(
tms, FLAGS.ol, sequences.PairWithIntermediates)
elif FLAGS.tds == 'tmb':
tms = [rpg_datasets_py.tum_mono.Sequence('%02d' % i)
for i in range(1, 51)]
return sequences.MultiSeqTrackingPairPicker(
tms, FLAGS.ol, sequences.PairWithIntermediates)
elif FLAGS.tds == 'tmbrc':
tms = [rpg_datasets_py.tum_mono.Sequence('%02d' % i)
for i in range(1, 51)] + \
[rpg_datasets_py.robotcar.CroppedGraySequence(
'2014-07-14-14-49-50')]
return sequences.MultiSeqTrackingPairPicker(
tms, FLAGS.ol, sequences.PairWithIntermediates)
elif FLAGS.tds == 'en':
ks = [rpg_datasets_py.kitti.KittiSeq(i)
for i in rpg_datasets_py.kitti.split('training')]
rs = rpg_datasets_py.robotcar.getSplitSequences('training')
tms = [rpg_datasets_py.tum_mono.Sequence(i)
for i in ['01', '02', '03', '48', '49', '50']]
pair_classes = [pair_class] * len(ks) + \
[sequences.PairWithIntermediates] * (len(rs) + len(tms))
return sequences.MultiSeqTrackingPairPicker(
ks + rs + tms, FLAGS.ol, pair_classes)
elif FLAGS.tds == 'hp':
data_gen = HPatches('training', use_min=True)
else:
assert False
return data_gen
def getEvalDataGen():
if not FLAGS.testing:
if FLAGS.ds == 'kt':
val_seqs = rpg_datasets_py.kitti.split('validation')
assert len(val_seqs) == 1
k = rpg_datasets_py.kitti.KittiSeq(val_seqs[0])
rpick = sequences.TrackingPairPicker(
k, 0.5, pair_class=sequences.PairWithStereo)
return sequences.FixPairs(rpick, 100)
else:
assert FLAGS.ds == 'hp'
ret = HPatches('validation', use_min=True)
print(ret.folder_names)
return ret
else:
if FLAGS.ds == 'hp':
ret = HPatches('testing', use_min=True)
print(ret.folder_names)
return ret
elif FLAGS.ds == 'kt':
val_seqs = rpg_datasets_py.kitti.split('testing')
assert len(val_seqs) == 1
k = rpg_datasets_py.kitti.KittiSeq(val_seqs[0])
rpick = sequences.TrackingPairPicker(
k, 0.5, pair_class=sequences.PairWithStereo)
return sequences.FixPairs(rpick, 100)
elif FLAGS.ds == 'eu':
seq = rpg_datasets_py.euroc.EurocSeq('V1_01_easy')
rpick = sequences.TrackingPairPicker(
seq, 0.5, pair_class=sequences.PairWithStereo)
return sequences.FixPairs(rpick, 100)
elif FLAGS.ds == 'eumh':
seq = rpg_datasets_py.euroc.EurocSeq('MH_01_easy')
rpick = sequences.TrackingPairPicker(
seq, 0.5, pair_class=sequences.PairWithStereo)
return sequences.FixPairs(rpick, 100)
else:
raise NotImplementedError
def getEvalSequences():
if FLAGS.ds == 'kt':
if FLAGS.testing:
seq_names = rpg_datasets_py.kitti.split('testing')
else:
seq_names = rpg_datasets_py.kitti.split('validation')
return [rpg_datasets_py.kitti.KittiSeq(i) for i in seq_names]
elif FLAGS.ds == 'eumh':
assert FLAGS.testing
seq_names = ['MH_01_easy']
return [rpg_datasets_py.euroc.EurocSeq(i) for i in seq_names]
else:
assert FLAGS.ds == 'eu'
assert FLAGS.testing
seq_names = ['V1_01_easy']
return [rpg_datasets_py.euroc.EurocSeq(i) for i in seq_names]
def getForwardPasser(graph=None, sess=None):
if FLAGS.baseline == '':
assert graph is not None
assert sess is not None
fp = system.ForwardPasser(
graph, sess, FLAGS.num_test_pts, FLAGS.nms)
if FLAGS.num_scales > 1:
return multiscale.ForwardPasser(
fp, FLAGS.scale_factor, FLAGS.num_scales)
else:
return fp
elif FLAGS.baseline in ['surf', 'sift']:
return baselines.OpenCVForwardPasser(FLAGS.baseline)
else:
raise Exception('Baseline %s unknown' % FLAGS.baseline)
def modelFromCheckpoint():
if FLAGS.baseline is not '':
return None, None
tf.reset_default_graph()
g = graph.Graph()
saver = tf.train.Saver()
sess = tf.Session()
saver.restore(sess, checkpointPath())
return g, sess
def checkpointRoot():
return os.path.join(os.path.dirname(__file__), 'checkpoints')
def checkpointPath():
if FLAGS.val_best:
return os.path.join(checkpointRoot(), shortString() + '_best')
else:
return os.path.join(checkpointRoot(), shortString())
def trainStatsPath():
return '%s_stats' % checkpointPath()
def methodString():
if FLAGS.baseline == '':
return shortString()
else:
return FLAGS.baseline
def evalString(k=True):
result = '%s_%d' % (FLAGS.ds, FLAGS.num_test_pts)
if k:
result += '_%d' % FLAGS.k
if FLAGS.testing:
result += '_TESTING'
return result
def vt():
if FLAGS.testing:
return 'testing'
else:
return 'validation'
def methodEvalString(k=True):
return '%s_%s' % (methodString(), evalString(k=k))
def announceEval():
announce('Evaluating %s...' % methodEvalString())
if FLAGS.testing:
announce('THIS IS A TESTING RUN!!!')
def resultPath(k=True):
return os.path.join('results', methodEvalString(k=k))
def wasInlierFilePath():
return '%s_wasinl' % resultPath()
def cachedForwardPath():
return '%s_cached_fp.hkl' % resultPath(k=False)
def evalPath():
return '%s_n_r_t.hkl' % resultPath()
def choutRootPath():
return '%s_chouts' % resultPath()
def seqFpsPath():
here = os.path.dirname(__file__)
return os.path.join(here, 'sequence_fps', methodEvalString(k=False))
def label():
if FLAGS.baseline != '':
return FLAGS.baseline
else:
if FLAGS.num_scales > 1:
return 'ours, multi-scale'
else:
return 'ours, single scale'
