def get(self, user_id, max_return):
try:
oe = os.environ
conn = psycopg2.connect(database=oe['DB_NAME'],
user=oe['DB_USER'],
password=oe['DB_PASSWORD'],
host=oe['DB_HOST'])
except Exception as e:
return str(e)
mc = Matcher(conn)
return json.dumps(mc.query(user_id, max_return))
def get(self, user_id, max_return):
try:
oe = os.environ
conn = psycopg2.connect(database=oe['DB_NAME'],
user=oe['DB_USER'],
password=oe['DB_PASSWORD'],
host=oe['DB_HOST'])
except Exception as e:
return str(e)
mc = Matcher(conn)
return json.dumps(mc.query(user_id, max_return))
def __init__(self, cfg):
self.cfg_ = self.build_cfg(cfg)
self.detector_ = cv2.FastFeatureDetector_create(threshold=19,
nonmaxSuppression=True)
self.extractor_ = cv2.ORB_create(2048, edgeThreshold=19)
#self.extractor_ = cv2.xfeatures2d.SURF_create()
self.matcher_ = Matcher(ex=self.extractor_)
self.tracker_ = Tracker(pLK=cfg['pLK'])
self.kf_ = build_ekf()
self.db_ = self.build_db()
self.state_ = PipelineState.INIT
# higher-level handles?
self.initializer_ = MapInitializer(db=self.build_db(),
matcher=self.matcher_,
tracker=self.tracker_,
cfg=self.cfg_)
def __init__(self,
gpu_id=0,
track_model=None,
pose_model=None,
embedding_model=None):
if self.tracker_flag:
self.tracker = SiamFCTracker(gpu_id, track_model)
self.posenet = PoseNet(gpu_id, pose_model)
self.matcher = Matcher()
print('----------------------------------------')
print('Flag parameters are set as follow:')
print('Tracker flag: {}'.format(self.tracker_flag))
print('Tracker update flag: {}'.format(self.tracker_update_flag))
print('Decrease tracker flag: {}'.format(self.descrease_tracker_flag))
print('New embedding(with pose) flag: {}'.format(
self.new_embedding_flag))
print('----------------------------------------')
def ontology_alignment(model, ontoTerms_a, ontoTerms_b, words, ceil = 0.5):
with open(ontoTerms_a) as f:
ontoText_a = f.readlines()
with open(ontoTerms_b) as f:
ontoText_b = f.readlines()
# Remove whitespace characters like `\n` at the end of each line.
ontoText_a = [x.strip() for x in ontoText_a]
ontoText_b = [x.strip() for x in ontoText_b]
whole = []
for text_a in ontoText_a:
for text_b in ontoText_b:
txt_a = re.sub(' +',' ',text_a)
txt_b = re.sub(' +',' ',text_b)
if txt_a == txt_b:
whole.append([text_a, text_b, 0.0])
try:
ontoText_a.remove(text_a)
except ValueError:
pass
#print(text_a)
try:
ontoText_b.remove(text_b)
except ValueError:
pass
#print(text_b)
# Transform to Word & Mask vectors to apply "feedforward_function"
ontoData_a, ontoData_b = [], []
for sentence in ontoText_a:
ontoData_a.append(getSeq(sentence, words))
for sentence in ontoText_b:
ontoData_b.append(getSeq(sentence, words))
x1,m1 = utils.prepare_data(ontoData_a)
x2,m2 = utils.prepare_data(ontoData_b)
OntoEmbg_a = model.feedforward_function(x1,m1)
OntoEmbg_b = model.feedforward_function(x2,m2)
# Compute the Cosine Distances:
dist = cosine_distances(OntoEmbg_a,OntoEmbg_b)
disT = np.transpose(dist)
males = preferances(dist)
females = preferances(disT)
del(disT)
match = Matcher(males, females)
marriage = match()
del(males); del(females)
for key, value in marriage.items():
man = ontoText_a[value]
woman = ontoText_b[key]
value = dist[value][key]
if value < ceil:
whole.append([man, woman, value])
return whole
def do_match():
print("Mentors:")
print(MR)
print("Mentee:")
print(ME)
# initialize Matcher with preference lists for both men and women
match = Matcher(MR, ME)
return match()
def __init__(
self,
gpu_id=[0, 0, 0, 0],
flag=[True, False, True, False],
#track_model='/export/home/zby/SiamFC/models/output/siamfc_35.pth',
track_model='/export/home/zby/SiamFC/models/output/siamfc_20.pth',
detection_model='/export/home/zby/SiamFC/models/res101_old/pascal_voc/faster_rcnn_1_25_4379.pth',
pose_model='/export/home/zby/SiamFC/data/models/final_new.pth.tar',
embedding_model='/export/home/zby/SiamFC/models/embedding_model.pth'
):
if flag[0]:
self.tracker = SiamFCTracker(gpu_id[0], track_model) #input RGB
if flag[1]:
self.detector = Detector(gpu_id[1], detection_model) #input BGR
if flag[2]:
self.posenet = PoseNet(gpu_id[2], pose_model) #input BGR
if flag[3]:
self.embedder = EmbeddingNet(gpu_id[3], embedding_model)
#self.tracker = SiamFCTracker(gpu_id[0], track_model)
self.matcher = Matcher()
print('----------------------------------------')
def __init__(
self,
gpu_id=0,
track_model='/export/home/zby/SiamFC/models/output/siamfc_35_old.pth',
#track_model = '/export/home/zby/SiamFC/data/models/siamfc_cpm_368_5.pth',
pose_model='/export/home/zby/SiamFC/data/models/final_new.pth.tar',
embedding_model='/export/home/zby/SiamFC/models/embedding_model.pth'
):
if self.tracker_flag:
self.tracker = SiamFCTracker(gpu_id, track_model)
self.posenet = PoseNet(gpu_id, pose_model)
if not self.new_embedding_flag:
self.embedder = EmbeddingNet(gpu_id, embedding_model)
self.matcher = Matcher()
print('----------------------------------------')
print('Flag parameters are set as follow:')
print('Tracker flag: {}'.format(self.tracker_flag))
print('Tracker update flag: {}'.format(self.tracker_update_flag))
print('Decrease tracker flag: {}'.format(self.descrease_tracker_flag))
print('New embedding(with pose) flag: {}'.format(
self.new_embedding_flag))
print('----------------------------------------')
def get(self, user_id, max_return, needs):
mc = Matcher()
# return json.dumps(mc.query(user_id, max_return, needs))
return json.dumps(mc.query(user_id, max_return, 'all'))
def main():
m = Matcher()
m.fit(TRAIN_FILE0, TRAIN_FILE1, TRAIN_MATCH_FILE)
m.transform(TEST_FILE0, TEST_FILE1)
m.write(OUTPUT_FILE)
def setup(v_, j_, flipped):
# the volunteers and their list of ordered job preferences
# v_ = dict((m, prefs.split(', ')) for [m, prefs] in (
# line.rstrip().split(': ') for line in open('volunteers.short.txt')))
# j_ = dict((m, prefs.split(', ')) for [m, prefs] in (
# line.rstrip().split(': ') for line in open('jobs.txt')))
volunteers = list(v_.keys())
jobs = list(j_.keys())
# print('type v_["abe"]', type(v_['abe']), volunteers)
# remove any missing job names from volunteers
for v in volunteers:
NA = v_[v][-1]
v_[v] = list(filter(lambda j: j in jobs, v_[v][:-1]))
v_[v].append(NA)
# remove any missing volunteer names from jobs
for j in jobs:
NA = j_[j][-1]
j_[j] = list(filter(lambda v: v in volunteers, j_[j][:-1]))
j_[j].append(NA)
J = {}
prefs = v_[list(v_.keys())[0]]
print('type(prefs)', type(prefs))
# prefs = prefs_.split(', ')
# NA = prefs[-1]
for p in jobs:
w__ = j_[p]
J[Person(p, int(w__[-1]))] = w__[:-1]
print(f'initial J keys {J.keys()}')
V = {}
prefs = j_[list(j_.keys())[0]]
# prefs = prefs_.split(', ')
for p in volunteers:
m__ = v_.get(p, ['0'])
person = Person(p, int(m__[-1]))
V[person] = []
for n in m__[:-1]:
# print('n', n)
job = list(filter(lambda j: j.n == n, J.keys()))[0]
V[person].append(job)
for j, prefs in J.items():
J[j] = []
for n in prefs:
volunteer = list(filter(lambda m: m.n == n, V.keys()))[0]
J[j].append(volunteer)
# for each volunteer construct a list of forbidden jobs
forbidden = {} # { 'dan': ['gay', 'eve', 'abi'], 'hal': ['eve'] }
for v, prefs in V.items():
NA = v.NA
# all jobs at or over the NA index are forbidden
forbidden[v] = prefs[NA:]
# n = random.randint(0, len(prefs) - 1)
# forbidden[m] = random.sample(prefs, n) # random sample of n wives
forbidden_v = {} # { 'dan': ['gay', 'eve', 'abi'], 'hal': ['eve'] }
for j, prefs in J.items():
NA = j.NA
# all volunteers at or over the NA index are forbidden
forbidden_v[j] = prefs[NA:]
C = defaultdict(list)
jKeys = set()
loop = 0
while len(J) > 0:
print("V & J")
print(V)
print(J)
match = Matcher(V, J, forbidden) # , forbidden_v)
# match volunteers and jobs; returns a mapping of jobs to volunteers
matches = match()
assert match.is_stable(matches) # should be a stable matching
print('stable match')
print(f'loop {loop} list(matches keys) {list(matches.keys())}')
loop += 1
# if loop > 2:
# break
# if len(C) == 0:
# C = dict((value, [key]) for key, value in enumerate(matches))
# print('Initial C.keys()', C.keys())
# else:
for _, key in enumerate(matches.items()):
C[key[1]].append(key[0])
print('Initial C.keys()', C.keys())
print('Initial C.values()', C.values())
jKeys |= set(matches.keys())
print(f"len jKeys {len(jKeys)} len(J) {len(J)} jKeys {jKeys} ")
J_ = copy.copy(J)
J = {}
for key, value in enumerate(J_.items()):
print(
f'J.items() key {key} value[0] {value[0]} type(value[0]) {type(value[0])} value[1] {value[1]}'
)
if value[0] in jKeys:
print(f'value {value[0]} in jKeys)')
else:
print(f'value {value[0]} NOT in jKeys)')
J[value[0]] = value[1]
print(f'len filtered J {len(J)} J {J}')
if len(J) == 0:
break
V_ = copy.copy(V)
for v, prefs in V_.items():
# print(f'k,v in V k {k} v {v}')
prefs = [p for p in prefs if p in list(J.keys())]
print(f'new prefs {prefs}')
V[v] = prefs
# V = {k: v for k, v in mydict.items() if k.startswith('foo')}
# J = dict((key, value) in enumerate(J.items()))
# J = dict((key, value) in enumerate(J.items()) if key not in jKeys)
# J = dict(filter(lamba j, v: j not in jKeys, enumerate(J))
# if len(J) > 0:
# print("len(J) > 0")
# V_ = sorted(C.items(), key=lambda kv: len(kv[1]))[:len(J)]
# match = Matcher(V_, J, forbidden, forbidden_v)
# # match volunteers and jobs; returns a mapping of jobs to volunteers
# matches = match()
# for key, value in enumerate(matches):
# C[value].append(key)
# print('jobs', jobs)
print('C.keys()', C.keys())
print([(key, value) for key, value in enumerate(C)])
if flipped:
a = [([j.n for j in value[1]], value[0].n)
for key, value in enumerate(C.items())]
else:
a = [(value[0].n, [j.n for j in value[1]])
for key, value in enumerate(C.items())]
# a = [(key.n, [j.n for j in C[key]]) for key in list(C.keys())]
# a=[(matches[key].n, key.n) for key in list(matches.keys())]
return jsonify(a)
from match import Matcher
# the men and their list of ordered spousal preferences
M = dict(
(m, prefs.split(', '))
for [m, prefs] in (line.rstrip().split(': ') for line in open('men.txt')))
# the women and their list of ordered spousal preferences
W = dict(
(m, prefs.split(', ')) for [m, prefs] in (line.rstrip().split(': ')
for line in open('women.txt')))
# initialize Matcher with preference lists for both men and women
match = Matcher(M, W)
# match men and women; returns a mapping of wives to husbands
wives = match()
assert match.is_stable() # should be a stable matching
# swap the husbands of two wives, which should make the matching unstable
wives['fay'], wives['gay'] = wives['gay'], wives['fay']
assert match.is_stable(wives) is False # should not be a stable matching
# with the perturbed matching we find that gav's marriage to fay is unstable:
#
# * gav prefers gay over fay
# * gay prefers gav over her current husband dan
class Pipeline(object):
def __init__(self, cfg):
self.cfg_ = self.build_cfg(cfg)
self.detector_ = cv2.FastFeatureDetector_create(threshold=19,
nonmaxSuppression=True)
self.extractor_ = cv2.ORB_create(2048, edgeThreshold=19)
#self.extractor_ = cv2.xfeatures2d.SURF_create()
self.matcher_ = Matcher(ex=self.extractor_)
self.tracker_ = Tracker(pLK=cfg['pLK'])
self.kf_ = build_ekf()
self.db_ = self.build_db()
self.state_ = PipelineState.INIT
# higher-level handles?
self.initializer_ = MapInitializer(db=self.build_db(),
matcher=self.matcher_,
tracker=self.tracker_,
cfg=self.cfg_)
def build_cfg(self, cfg):
# build derived values
# apply scale
w = int(cfg['scale'] * cfg['w'])
h = int(cfg['scale'] * cfg['h'])
K0 = cfg['K']
K = cfg['scale'] * cfg['K']
K[2, 2] = 1.0
# image shape
shape = (h, w, 3) # TODO : handle monochrome
# first, make a copy from argument
cfg = dict(cfg)
# insert derived values
cfg['w'] = w
cfg['h'] = h
cfg['shape'] = shape
cfg['K0'] = K0
cfg['K'] = K
# create dynamic type
#ks = cfg.keys()
#cfg_t = namedtuple('PipelineConfig', ks)
# setup dot-referenced aliases
# for k, v in cfg.iteritems():
# setattr(cfg, k, v)
return cfg
def build_db(self):
cfg = self.cfg_
ex = self.extractor_
img_fmt = (cfg['shape'], np.uint8)
dsc_t = (np.uint8 if ex.descriptorType() == cv2.CV_8U else np.float32)
dsc_fmt = (self.extractor_.descriptorSize(), dsc_t)
return DB(img_fmt=img_fmt, dsc_fmt=dsc_fmt)
def motion_model(self, f0, f1, stamp, use_kalman=False):
if not use_kalman:
# simple `repetition` model
txn0, rxn0 = f0['pose'][L_POS], f0['pose'][A_POS]
txn1, rxn1 = f1['pose'][L_POS], f1['pose'][A_POS]
R0 = tx.euler_matrix(*rxn0)
R1 = tx.euler_matrix(*rxn1)
T0 = tx.compose_matrix(angles=rxn0, translate=txn0)
T1 = tx.compose_matrix(angles=rxn1, translate=txn1)
Tv = np.dot(T1, vm.inv(T0)) # Tv * T0 = T1
T2 = np.dot(Tv, T1)
txn = tx.translation_from_matrix(T2)
rxn = tx.euler_from_matrix(T2)
x = f1['pose'].copy()
P = f1['cov'].copy()
x[0:3] = txn
x[9:12] = rxn
return x, P
else:
# dt MUST NOT BE None
self.kf_.x = f0['pose']
self.kf_.P = f0['cov']
dt = (f1['stamp'] - f0['stamp'])
self.kf_.predict(dt)
txn, rxn = f1['pose'][L_POS], f1['pose'][A_POS]
z = np.concatenate([txn, rxn])
self.kf_.update(z)
dt = (stamp - f1['stamp'])
self.kf_.predict(dt)
return self.kf_.x.copy(), self.kf_.P.copy()
def is_keyframe(self, frame):
# TODO : more robust keyframe heuristic
# == possibly filter for richness of tracking features?
feat = (frame['feat']).item()
return len(feat.kpt) > 100 # TODO: arbitrary threshold
def build_frame(self, img, stamp):
""" build a simple frame """
# automatic index assignment
# NOTE: multiple frames will have the same index
# if not appended to self.db_.frame
# TODO : separate out feature processing parts?
index = self.db_.frame.size
# by default, not a keyframe
is_kf = False
# extract features
kpt = self.detector_.detect(img)
kpt, dsc = self.extractor_.compute(img, kpt)
# kpt, dsc = self.extractor_.detectAndCompute(img, None)
feat = Feature(kpt, dsc, cv2.KeyPoint.convert(kpt))
# apply motion model? initialize pose anyway
if self.db_.frame_.size >= 2:
print('motion-model')
x, P = self.motion_model(f0=self.db_.frame_[-2],
f1=self.db_.frame_[-1],
stamp=stamp,
use_kalman=True)
else:
x = np.zeros(self.cfg_['state_size'])
P = 1e-6 * np.eye(self.cfg_['state_size'])
frame = (index, stamp, img, x, P, is_kf, feat)
res = np.array(frame, dtype=self.db_.frame.dtype)
return res
def transition(self, new_state):
print('[state] ({} -> {})'.format(self.state_, new_state))
self.state_ = new_state
def init_map(self, img, stamp, data):
""" initialize map """
# fetch prv+cur frames
# populate frame from motion model
frame = self.build_frame(img, stamp)
suc = self.initializer_.compute(frame, data)
# print(data['dbg-tv'])
if not suc:
return
self.db_.extend(self.initializer_.db_)
self.transition(PipelineState.TRACK)
#print self.db_.landmark_['pos'][self.db_.landmark_['tri']][:5]
#print self.initializer_.db_.landmark_['pos'][self.db_.landmark_['tri']][:5]
if True:
frame0 = self.db_.frame_[0]
frame1 = self.db_.frame_[1]
img0, img1 = frame0['image'], frame1['image']
feat0, feat1 = frame0['feat'], frame1['feat']
pt0m, pt1m = feat0.pt[data['mi0']], feat1.pt[data['mi1']]
msk = data['msk_cld']
print('frame pair : {}-{}'.format(frame0['index'],
frame1['index']))
viz0 = cv2.drawKeypoints(img0, feat0.kpt, None)
viz1 = cv2.drawKeypoints(img1, feat1.kpt, None)
viz = draw_matches(viz0, viz1, pt0m[msk], pt1m[msk])
data['viz'] = viz
# == if cfg['dbg-cloud']:
dr_data = {}
cld_viz, col_viz = DenseRec(self.cfg_['K']).compute(img0,
img1,
P1=data['P0'],
P2=data['P1'],
data=dr_data)
cdist = vm.norm(cld_viz)
data['cld_viz'] = cld_viz[cdist < np.percentile(cdist, 95)]
data['col_viz'] = col_viz[cdist < np.percentile(cdist, 95)]
self.initializer_.reset()
return
def bundle_adjust(self, frame0, frame1):
idx0, idx1 = max(frame0['index'], frame1['index'] - 8), frame1['index']
#idx0, idx1 = keyframe['index'], frame1['index']
obs = self.db_.observation
msk = np.logical_and(idx0 <= obs['src_idx'], obs['src_idx'] <= idx1)
# parse observation
i_src = obs['src_idx'][msk]
#print('i_src', i_src)
i_lmk = obs['lmk_idx'][msk]
p_obs = obs['point'][msk]
# index pruning relevant sources
i_src_alt, i_a2s, i_s2a = index_remap(i_src)
i_lmk_alt, i_a2l, i_l2a = index_remap(i_lmk)
# 1. select targets based on new index
i_src = i_s2a
i_lmk = i_l2a
frames = self.db_.frame[i_a2s[i_src_alt]]
landmarks = self.db_.landmark[i_a2l[i_lmk_alt]]
# parse data
txn = frames['pose'][:, L_POS]
rxn = frames['pose'][:, A_POS]
lmk = landmarks['pos']
data_ba = {}
# NOTE : txn/rxn will be internally inverted to reduce duplicate compute.
suc = BundleAdjustment(
i_src,
i_lmk,
p_obs, # << observation
txn,
rxn,
lmk,
self.cfg_['K'],
axa=True).compute(data=data_ba) # << data
if suc:
# TODO : apply post-processing kalman filter?
#print('{}->{}'.format(txn, data_ba['txn']))
#print('{}->{}'.format(rxn, data_ba['rxn']))
txn = data_ba['txn']
rxn = data_ba['rxn']
lmk = data_ba['lmk']
self.db_.frame['pose'][i_a2s[i_src_alt], L_POS] = txn
self.db_.frame['pose'][i_a2s[i_src_alt], A_POS] = rxn
self.db_.landmark['pos'][i_a2l[i_lmk_alt]] = lmk
@profile(sort='cumtime')
def track(self, img, stamp, data={}):
""" Track landmarks"""
# unroll data
# fetch frame pair
# TODO : add landmarks along the way
# TODO : update landmarks through optimization
mapframe = self.db_.keyframe[0] # first keyframe = map frame
keyframe = self.db_.keyframe[-1] # last **keyframe**
frame0 = self.db_.frame[-1] # last **frame**
frame1 = self.build_frame(img, stamp)
# print('prior position',
# frame1['pose'][L_POS], frame1['pose'][A_POS])
landmark = self.db_.landmark
img1 = frame1['image']
feat1 = frame1['feat'].item()
# bypass match_local for already tracking points ...
pt0_l = landmark['pt'][landmark['track']]
pt1_l, msk_t = self.tracker_.track(frame0['image'],
img1,
pt0_l,
return_msk=True)
# apply tracking mask
pt0_l = pt0_l[msk_t]
pt1_l = pt1_l[msk_t]
# update tracking status
landmark['track'][landmark['track'].nonzero()[0][~msk_t]] = False
landmark['pt'][landmark['track']] = pt1_l
# search additional points
cld0_l = landmark['pos'][~landmark['track']]
dsc_l = landmark['dsc'][~landmark['track']]
msk_prj = None
if len(cld0_l) >= 128:
# merge with projections
pt0_cld_l = project_to_frame(cld0_l,
source_frame=mapframe,
target_frame=frame1,
K=self.cfg_['K'],
D=self.cfg_['D'])
# in-frame projection mask
msk_prj = np.logical_and.reduce([
0 <= pt0_cld_l[..., 0],
pt0_cld_l[..., 0] < self.cfg_['w'],
0 <= pt0_cld_l[..., 1],
pt0_cld_l[..., 1] < self.cfg_['h'],
])
pt0 = pt0_l
pt1 = pt1_l
cld0 = landmark['pos'][landmark['track']]
obs_lmk_idx = landmark['index'][landmark['track']]
search_map = (False and len(cld0) <= 256 and (msk_prj is not None)
and (msk_prj.sum() >= 16))
if search_map:
# sample points from the map
mi0, mi1 = match_local(
pt0_cld_l[msk_prj],
feat1.pt,
dsc_l[msk_prj],
feat1.dsc,
hamming=(not feat1.dsc.dtype == np.float32),
)
# collect all parts
pt0 = np.concatenate([pt0, pt0_cld_l[msk_prj][mi0]], axis=0)
pt1 = np.concatenate([pt1, feat1.pt[mi1]], axis=0)
cld0 = np.concatenate(
[cld0, landmark['pos'][~landmark['track']][msk_prj][mi0]],
axis=0)
obs_lmk_idx = np.concatenate([
obs_lmk_idx,
landmark['index'][~landmark['track']][msk_prj][mi0]
],
axis=0)
# debug ...
# pt_dbg = project_to_frame(
# landmark['pos'][landmark['track']],
# frame1,
# self.cfg_['K'], self.cfg_['D'])
##img_dbg = draw_points(img1.copy(), pt_dbg, color=(255,0,0) )
##draw_points(img_dbg, pt1, color=(0,0,255) )
#img_dbg = draw_matches(img1, img1, pt_dbg, pt1)
#cv2.imshow('dbg', img_dbg)
#print_ratio(len(pt0_l), len(pt0), name='point source')
# if len(mi0) <= 0:
# viz1 = draw_points(img1.copy(), pt0)
# viz2 = draw_points(img1.copy(), feat1.pt)
# viz = np.concatenate([viz1, viz2], axis=1)
# cv2.imshow('pnp', viz)
# return False
#print_ratio(len(mi0), len(pt0))
# suc, rvec, tvec = cv2.solvePnP(
# cld0[:, None], pt1[:, None],
# self.cfg_['K'], self.cfg_['D'],
# flags = cv2.SOLVEPNP_EPNP
# ) # T(rv,tv) . cld = cam
#inl = None
#print 'euler', tx.euler_from_matrix(cv2.Rodrigues(rvec)[0])
T_i = tx.inverse_matrix(
tx.compose_matrix(translate=frame1['pose'][L_POS],
angles=frame1['pose'][A_POS]))
rvec0 = cv2.Rodrigues(T_i[:3, :3])[0]
tvec0 = T_i[:3, 3:]
if len(pt1) >= 1024:
# prune
nmx_idx = non_max(pt1, landmark['rsp'][obs_lmk_idx])
print_ratio(len(nmx_idx), len(pt1), name='non_max')
cld_pnp, pt1_pnp = cld0[nmx_idx], pt1[nmx_idx]
else:
cld_pnp, pt1_pnp = cld0, pt1
# hmm ... pose-only BA vs. PnP
if False:
data_pnp = {}
#print('txn-prv', frame0['pose'][L_POS])
#print('rxn-prv', frame0['pose'][A_POS])
#print('txn-in', frame1['pose'][L_POS])
#print('rxn-in', frame1['pose'][A_POS])
suc = BundleAdjustment(i_src=np.full(len(cld_pnp), 0),
i_lmk=np.arange(len(cld_pnp)),
p_obs=pt1_pnp,
txn=frame1['pose'][L_POS][None, ...],
rxn=frame1['pose'][A_POS][None, ...],
lmk=cld_pnp,
K=self.cfg_['K'],
pose_only=True).compute(crit=dict(
loss='soft_l1',
xtol=1e-8,
f_scale=np.sqrt(5.991)),
data=data_pnp)
#print('txn-out', data_pnp['txn'][0])
#print('rxn-out', data_pnp['rxn'][0])
T = tx.compose_matrix(translate=data_pnp['txn'][0],
angles=data_pnp['rxn'][0])
Ti = tx.inverse_matrix(T)
rxn_pnp = np.float32(tx.euler_from_matrix(Ti))
txn_pnp = tx.translation_from_matrix(Ti)
rvec = cv2.Rodrigues(Ti[:3, :3])[0]
tvec = txn_pnp
prj = cvu.project_points(cld_pnp, rvec, tvec, self.cfg_['K'],
self.cfg_['D'])
err = vm.norm(prj - pt1_pnp)
inl = np.where(err <= 1.0)[0]
else:
# == if(cfg['dbg_pnp']):
#print 'frame1-pose', frame1['pose']
# dbg = draw_matches(img1, img1,
# project_to_frame(cld_pnp, source_frame=mapframe, target_frame=frame1,
# K=self.cfg_['K'], D=self.cfg_['D']),
# pt1_pnp)
#cv2.imshow('pnp', dbg)
# cv2.waitKey(0)
suc, rvec, tvec, inl = cv2.solvePnPRansac(
cld_pnp[:, None],
pt1_pnp[:, None],
self.cfg_['K'],
self.cfg_['D'],
useExtrinsicGuess=True,
rvec=rvec0,
tvec=tvec0,
iterationsCount=1024,
reprojectionError=1.0,
confidence=0.999,
flags=cv2.SOLVEPNP_EPNP
# flags=cv2.SOLVEPNP_DLS
# flags=cv2.SOLVEPNP_ITERATIVE
# minInliersCount=0.5*_['pt0']
)
n_pnp_in = len(cld_pnp)
n_pnp_out = len(inl) if (inl is not None) else 0
#print 'inl', inl
print n_pnp_in
print n_pnp_out
suc = (suc and (inl is not None)
and (n_pnp_out >= 128 or n_pnp_out >= 0.25 * n_pnp_in))
print('pnp success : {}'.format(suc))
if inl is not None:
print_ratio(n_pnp_out, n_pnp_in, name='pnp')
# visualize match statistics
viz_pt0 = project_to_frame(
cld0,
source_frame=mapframe,
target_frame=keyframe, # TODO: keyframe may no longer be true?
K=self.cfg_['K'],
D=self.cfg_['D'])
viz_msk = np.logical_and.reduce([
0 <= viz_pt0[:, 0],
viz_pt0[:, 0] < self.cfg_['w'],
0 <= viz_pt0[:, 1],
viz_pt0[:, 1] < self.cfg_['h'],
])
viz1 = draw_points(img1.copy(), feat1.pt)
viz = draw_matches(keyframe['image'], viz1, viz_pt0[viz_msk],
pt1[viz_msk])
data['viz'] = viz
# obtained position!
R = cv2.Rodrigues(rvec)[0]
t = np.float32(tvec)
R, t = vm.Rti(R, t)
rxn = np.reshape(tx.euler_from_matrix(R), 3)
txn = t.ravel()
if suc:
# print('pnp-txn', t)
# print('pnp-rxn', tx.euler_from_matrix(R))
# motion_update()
if self.cfg_['kalman']:
# kalman_update()
self.kf_.x = frame0['pose']
self.kf_.P = frame0['cov']
self.kf_.predict(frame1['stamp'] - frame0['stamp'])
self.kf_.update(np.concatenate([txn, rxn]))
frame1['pose'] = self.kf_.x
frame1['cov'] = self.kf_.P
else:
# hard_update()
frame1['pose'][L_POS] = t.ravel()
frame1['pose'][A_POS] = tx.euler_from_matrix(R)
self.db_.observation.extend(
dict(
# observation frame source
src_idx=np.full_like(obs_lmk_idx, frame1['index']),
lmk_idx=obs_lmk_idx, # landmark index
point=pt1))
self.db_.frame.append(frame1)
x = 1
need_kf = np.logical_or.reduce([
not suc, # PNP failed -- try new keyframe
# PNP was decent but would be better to have a new frame
suc and (n_pnp_out < 128),
# = frame is somewhat stale
(frame1['index'] - keyframe['index'] > 32) and (msk_t.sum() < 256)
]) and self.is_keyframe(frame1)
#need_kf = False
# ?? better criteria for running BA?
run_ba = (frame1['index'] % 8) == 0
#run_ba = False
#run_ba = need_kf
if run_ba:
self.bundle_adjust(keyframe, frame1)
if need_kf:
for index in reversed(range(keyframe['index'], frame1['index'])):
feat0, feat1 = self.db_.frame[index]['feat'], frame1[
'feat'].item()
mi0, mi1 = self.matcher_.match(feat0.dsc,
feat1.dsc,
lowe=0.8,
fold=False)
data_tv = {}
suc_tv, det_tv = TwoView(feat0.pt[mi0], feat1.pt[mi1],
self.cfg_['K']).compute(data=data_tv)
if suc_tv:
print('======================= NEW KEYFRAME ===')
xfm0 = pose_to_xfm(self.db_.frame[index]['pose'])
xfm1 = pose_to_xfm(frame1['pose'])
scale_ref = np.linalg.norm(
tx.translation_from_matrix(vm.Ti(xfm1).dot(xfm0)))
scale_tv = np.linalg.norm(data_tv['t'])
# TODO : does NOT consider "duplicate" landmark identities
# IMPORTANT: frame1 is a `copy` of "last_frame"
#frame1['is_kf'] = True
self.db_.frame[-1]['is_kf'] = True
lmk_idx0 = self.db_.landmark.size
print('lmk_idx0', lmk_idx0)
msk_cld = data_tv['msk_cld']
cld1 = data_tv['cld1'][msk_cld] * (scale_ref / scale_tv)
cld = transform_cloud(
cld1,
source_frame=frame1,
target_frame=mapframe,
)
col = extract_color(frame1['image'],
feat1.pt[mi1][msk_cld])
local_map = dict(
index=lmk_idx0 + np.arange(len(cld)), # landmark index
src=np.full(len(cld), frame1['index']), # source index
dsc=feat1.dsc[mi1][msk_cld], # landmark descriptor
rsp=[
feat1.kpt[i].response
for i in np.arange(len(feat1.kpt))[mi1][msk_cld]
], # response "strength"
# tracking point initialization
pt0=feat1.pt[mi1][msk_cld],
invd=1.0 / cld1[..., 2],
depth=cld1[..., 2],
pos=cld, # landmark position [[ map frame ]]
# tracking status
track=np.ones(len(cld), dtype=np.bool),
# tracking point initialization
pt=feat1.pt[mi1][msk_cld],
tri=np.ones(len(cld), dtype=np.bool),
col=col, # debug : point color information
)
# hmm?
self.db_.landmark.extend(local_map)
break
else:
print('Attempted new keyframe but failed')
def process(self, img, stamp, data={}):
if self.state_ == PipelineState.IDLE:
return
# if self.state_ == PipelineState.NEED_REF:
# return self.init_ref(img, stamp, data)
# elif self.state_ == PipelineState.NEED_MAP:
# return self.init_map(img, stamp, data)
if self.state_ == PipelineState.INIT:
return self.init_map(img, stamp,
data) # self.initializer_.compute(
elif self.state_ == PipelineState.TRACK:
return self.track(img, stamp, data)
def save(self, path):
if not os.path.exists(path):
os.makedirs(path)
def D_(p):
return os.path.join(path, p)
np.save(D_('config.npy'), self.cfg_)
self.db_.save(path)
from match import Matcher
# the men and their list of ordered spousal preferences
M = dict((m, prefs.split(", ")) for [m, prefs] in (line.rstrip().split(": ") for line in open("men.txt")))
# the women and their list of ordered spousal preferences
W = dict((m, prefs.split(", ")) for [m, prefs] in (line.rstrip().split(": ") for line in open("women.txt")))
# initialize Matcher with preference lists for both men and women
match = Matcher(M, W)
# match men and women; returns a mapping of wives to husbands
wives = match()
assert match.is_stable() # should be a stable matching
# swap the husbands of two wives, which should make the matching unstable
wives["fay"], wives["gay"] = wives["gay"], wives["fay"]
assert match.is_stable(wives) is False # should not be a stable matching
# with the perturbed matching we find that gav's marriage to fay is unstable:
#
# * gav prefers gay over fay
# * gay prefers gav over her current husband dan
class Track_And_Detect:
effective_track_thresh = 0.55
effective_detection_thresh = 0.4
effective_keypoints_thresh = 0.6
effective_keypoints_number = 8
iou_match_thresh = 0.5
embedding_match_thresh = 2
nms_thresh = 0.5
oks_thresh = 0.8
def __init__(
self,
gpu_id=[0, 0, 0, 0],
flag=[True, False, True, False],
#track_model='/export/home/zby/SiamFC/models/output/siamfc_35.pth',
track_model='/export/home/zby/SiamFC/models/output/siamfc_20.pth',
detection_model='/export/home/zby/SiamFC/models/res101_old/pascal_voc/faster_rcnn_1_25_4379.pth',
pose_model='/export/home/zby/SiamFC/data/models/final_new.pth.tar',
embedding_model='/export/home/zby/SiamFC/models/embedding_model.pth'
):
if flag[0]:
self.tracker = SiamFCTracker(gpu_id[0], track_model) #input RGB
if flag[1]:
self.detector = Detector(gpu_id[1], detection_model) #input BGR
if flag[2]:
self.posenet = PoseNet(gpu_id[2], pose_model) #input BGR
if flag[3]:
self.embedder = EmbeddingNet(gpu_id[3], embedding_model)
#self.tracker = SiamFCTracker(gpu_id[0], track_model)
self.matcher = Matcher()
print('----------------------------------------')
#initialize the first frame of this video
def init_tracker(self, frame, bbox_list):
self.new_id_flag = 0
self.track_id_dict = dict()
#self.tracker.clear_data()
#conver bgr(opencv) to rgb
rgb_frame = frame[:, :, ::-1]
bbox_list, keypoint_list = self.oks_filter(bbox_list, frame)
#print(bbox_list)
for bbox in bbox_list:
self.create_id(frame, rgb_frame, bbox)
bbox_list = []
#pose_list=[]
for id, item in self.track_id_dict.items():
bbox = item['bbox_and_score'] + [id]
bbox_list.append(bbox)
# pose_position, pose_value, pose_heatmap = self.pose_detect(frame, bbox)
# pose_info = np.hstack(pose_postion, pose_value)
return bbox_list
def oks_filter(self, det_list, frame):
keypoint_list = []
for bbox in det_list:
center, scale = self.posenet.x1y1x2y2_to_cs(bbox[0:4])
area = np.prod(scale * 200, 1)
pred = np.zeros((15, 3), dtype=np.float32)
pose_positions, pose_vals, pose_heatmaps = self.pose_detect(
frame, bbox)
#print(pose_vals)
#posa_vals = np.expand_dims(pose_vals, axis=1)
pred[:, 0:2] = pose_positions
pred[:, 2] = pose_vals
score_all, valid_num = 0, 0
for i in range(15):
score_i = pose_vals[i]
if score_i >= 0.2:
score_all += score_i
valid_num += 1
if valid_num != 0:
new_score = score_all / valid_num * bbox[4]
else:
new_score = 0
keypoint_dict = {
'score': new_score,
'area': area,
'keypoints': pred
}
keypoint_list.append(keypoint_dict)
keep = self.matcher.oks_nms(keypoint_list, thresh=self.oks_thresh)
new_det_list = [det_list[i] for i in keep]
new_keypoint_list = [keypoint_list[i] for i in keep]
return new_det_list, new_keypoint_list
def create_id(self, frame, rgb_frame, bbox):
score = bbox[4]
bbox = bbox[0:4]
track_id = self.new_id_flag
feature = self.embedding(frame, bbox)
#self.track_id_dict[track_id]={'bbox_and_score':bbox+[score]}
self.track_id_dict[track_id] = {
'bbox_and_score': bbox + [score],
'feature': feature,
'frame_flag': 1,
'exist': True
}
#self.track_id_dict[track_id]={'bbox_and_score':bbox+[score],'feature':[feature], 'frame_flag':1, 'exist':True}
#pose_position, pose_value, pose_heatmap = self.pose_detect(frame, bbox)
#self.update_tracker(rgb_frame, bbox, track_id)
#print('Track id {} has been initinized'.format(track_id))
self.new_id_flag += 1
def update_id(self, frame, rgb_frame, det_bbox, track_id):
bbox, score = det_bbox[0:4], det_bbox[4]
feature = np.array(self.embedding(frame, bbox))
former_track_dict = self.track_id_dict[track_id]
former_frame_flag, former_feature = former_track_dict[
'frame_flag'], np.array(former_track_dict['feature'])
now_frame_flag = former_frame_flag + 1
#calculate the average feature
#now_feature = ((former_feature*former_frame_flag+feature)/now_frame_flag).tolist()
#former_feature = former_feature.tolist()
now_feature = feature.tolist()
#former_feature.append(now_feature)
self.track_id_dict[track_id] = {
'bbox_and_score': det_bbox,
'feature': now_feature,
'frame_flag': now_frame_flag,
'exist': True
}
#pose_position, pose_value, pose_heatmap = self.pose_detect(frame, bbox)
#self.update_tracker(rgb_frame, bbox, track_id)
def multi_track(self, frame):
rgb_frame = frame[:, :, ::-1]
bbox_list = []
for id in self.track_id_dict:
if self.track_id_dict[id]['exist'] == False:
continue
bbox, score = self.tracker.track_id(rgb_frame, id)
bbox_list.append(bbox + [score] + [id])
self.track_id_dict[id]['bbox'] = bbox
return bbox_list
def match_and_track_embedding(self, detections, track_list, frame):
# print(self.track_id_dict.keys())
# exist_flag = []
# for id in self.track_id_dict:
# exist_flag.append(self.track_id_dict[id]['exist'])
# print(exist_flag)
rgb_frame = frame[:, :, ::-1]
matches, unmatched_detections, unmatched_trackers = self.matcher.associate_detections_to_trackers_iou(
detections, track_list, iou_threshold=self.iou_match_thresh)
has_tracked_id = set()
for match in matches:
det_index, track_index = match
has_tracked_id.add(track_index)
det_bbox = detections[det_index]
update_id = track_list[track_index][5]
self.update_id(frame, rgb_frame, det_bbox, update_id)
#get feature unmatched_detections
det_feature_list = []
for new_index in unmatched_detections:
det_bbox = detections[new_index]
det_score = det_bbox[4]
det_id_feature = self.embedding(frame, det_bbox) + det_bbox
det_feature_list.append(det_id_feature)
#get feature for unmatched_trackers
database_feature_list = []
for database_id in self.track_id_dict:
if database_id in has_tracked_id:
continue
database_id_feature = self.track_id_dict[database_id][
'feature'] + [database_id]
database_feature_list.append(database_id_feature)
# for delete_index in unmatched_trackers:
# track_bbox = track_list[delete_index]
# track_score, delete_id = track_bbox[4], track_bbox[5]
# delete_id_feature = self.track_id_dict[delete_id]['feature'] + [delete_id]
# track_feature_list.append(delete_id_feature)
#match the detection and tracklist
embedding_matches, \
embedding_unmatched_detections,\
embedding_unmatched_trackers = self.matcher.associate_detections_to_trackers_embedding(det_feature_list,
database_feature_list,
distance_threshold = self.embedding_match_thresh)
for match in embedding_matches:
det_index, track_index = match
det_bbox = det_feature_list[det_index][2048:]
update_id = database_feature_list[track_index][2048]
self.update_id(frame, rgb_frame, det_bbox, update_id)
for new_index in embedding_unmatched_detections:
det_bbox = det_feature_list[new_index][2048:]
det_score = det_bbox[4]
pose_position, pose_value, pose_heatmap = self.pose_detect(
frame, det_bbox)
if det_score >= self.effective_detection_thresh and np.sum(
pose_value >= self.effective_keypoints_thresh
) >= self.effective_keypoints_number:
self.create_id(frame, rgb_frame, det_bbox)
#change status for unmatched_trackers
for delete_index in embedding_unmatched_trackers:
delete_id = track_feature_list[delete_index][2048]
self.track_id_dict[delete_id]['exist'] = False
#delete unuseful index for unmatched_trackers
# for delete_index in embedding_unmatched_trackers:
# delete_id = track_feature_list[delete_index][2048]
# del self.track_id_dict[delete_id]
# self.tracker.delete_id(delete_id)
bbox_list = []
for id, item in self.track_id_dict.items():
if item['exist'] == True:
bbox_list.append(item['bbox_and_score'] + [id])
#print(bbox_list)
return bbox_list
def match_detection_embedding(self, detections, frame):
rgb_frame = frame[:, :, ::-1]
#get detection feature
det_feature_list = []
for det_bbox in detections:
det_id_feature = self.embedding(frame, det_bbox) + det_bbox
det_feature_list.append(det_id_feature)
#get feature for former trackers
database_feature_list = []
for database_id in self.track_id_dict:
database_id_feature = self.track_id_dict[database_id][
'feature'] + [database_id]
database_feature_list.append(database_id_feature)
#match the detection and tracklist
embedding_matches, \
embedding_unmatched_detections,\
embedding_unmatched_trackers = self.matcher.associate_detections_to_trackers_embedding(det_feature_list,
database_feature_list,
distance_threshold = self.embedding_match_thresh)
for match in embedding_matches:
det_index, track_index = match
det_bbox = det_feature_list[det_index][2048:]
update_id = database_feature_list[track_index][2048]
self.update_id(frame, rgb_frame, det_bbox, update_id)
for new_index in embedding_unmatched_detections:
det_bbox = det_feature_list[new_index][2048:]
det_score = det_bbox[4]
pose_position, pose_value, pose_heatmap = self.pose_detect(
frame, det_bbox)
if det_score >= self.effective_detection_thresh and np.sum(
pose_value >= self.effective_keypoints_thresh
) >= self.effective_keypoints_number:
self.create_id(frame, rgb_frame, det_bbox)
#delete unuseful trackers
for delete_index in embedding_unmatched_trackers:
delete_id = database_feature_list[delete_index][2048]
del self.track_id_dict[delete_id]
self.tracker.delete_id(delete_id)
bbox_list = []
for id, item in self.track_id_dict.items():
if item['exist'] == True:
bbox_list.append(item['bbox_and_score'] + [id])
#print(bbox_list)
return bbox_list
def match_detection_iou(self, detections, frame):
rgb_frame = frame[:, :, ::-1]
#get feature for former trackers
database_bbox_list = []
for database_id in self.track_id_dict:
database_id_bbox = self.track_id_dict[database_id][
'bbox_and_score'] + [database_id]
database_bbox_list.append(database_id_bbox)
matches, unmatched_detections, unmatched_trackers = self.matcher.associate_detections_to_trackers_iou(
detections,
database_bbox_list,
iou_threshold=self.iou_match_thresh)
#update the matched trackers with detection bbox
for match in matches:
det_index, track_index = match
det_bbox = detections[det_index]
update_id = database_bbox_list[track_index][5]
self.track_id_dict[update_id]['bbox_and_score'] = det_bbox[0:5]
#create new index for unmatched_detections
for new_index in unmatched_detections:
det_bbox = detections[new_index]
det_score = det_bbox[4]
pose_position, pose_value, pose_heatmap = self.pose_detect(
frame, det_bbox)
if det_score >= self.effective_detection_thresh and np.sum(
pose_value >= self.effective_keypoints_thresh
) >= self.effective_keypoints_number:
self.create_id(frame, rgb_frame, det_bbox)
#delete unuseful index for unmatched_trackers
for delete_index in unmatched_trackers:
delete_id = database_bbox_list[delete_index][5]
del self.track_id_dict[delete_id]
self.tracker.delete_id(delete_id)
bbox_list = []
for id, item in self.track_id_dict.items():
if item['exist'] == True:
bbox_list.append(item['bbox_and_score'] + [id])
#print(bbox_list)
return bbox_list
def match_detection_iou_embedding(self, detections, frame):
rgb_frame = frame[:, :, ::-1]
#final_bbox = []
detections, keypoint_list = self.oks_filter(detections, frame)
#get feature for former trackers
database_bbox_list = []
for database_id in self.track_id_dict:
database_id_bbox = self.track_id_dict[database_id][
'bbox_and_score'] + [database_id]
database_bbox_list.append(database_id_bbox)
matches, unmatched_detections, unmatched_trackers = self.matcher.associate_detections_to_trackers_iou(
detections,
database_bbox_list,
iou_threshold=self.iou_match_thresh)
#update the matched trackers with detection bbox
for match in matches:
det_index, track_index = match
det_bbox = detections[det_index]
update_id = database_bbox_list[track_index][5]
self.update_id(frame, rgb_frame, det_bbox, update_id)
#final_bbox.append(det_bbox+[update_id])
#create new index for unmatched_detections
det_feature_list = []
for new_index in unmatched_detections:
det_bbox = detections[new_index]
det_score = det_bbox[4]
det_id_feature = self.embedding(frame, det_bbox) + det_bbox
det_feature_list.append(det_id_feature)
track_feature_list = []
for delete_index in unmatched_trackers:
track_bbox = database_bbox_list[delete_index]
track_score, delete_id = track_bbox[4], track_bbox[5]
delete_id_feature = self.track_id_dict[delete_id]['feature'] + [
delete_id
]
track_feature_list.append(delete_id_feature)
embedding_matches, \
embedding_unmatched_detections,\
embedding_unmatched_trackers = self.matcher.associate_detections_to_trackers_embedding(det_feature_list,
track_feature_list,
distance_threshold = self.embedding_match_thresh)
#update matched embedding detections and former tracking feature
for match in embedding_matches:
det_index, track_index = match
det_bbox = det_feature_list[det_index][2048:]
update_id = track_feature_list[track_index][2048]
self.update_id(frame, rgb_frame, det_bbox, update_id)
#create new id for unmatched detections
for new_index in embedding_unmatched_detections:
det_bbox = det_feature_list[new_index][2048:]
det_score = det_bbox[4]
pose_position, pose_value, pose_heatmap = self.pose_detect(
frame, det_bbox)
if det_score >= self.effective_detection_thresh and np.sum(
pose_value >= self.effective_keypoints_thresh
) >= self.effective_keypoints_number:
self.create_id(frame, rgb_frame, det_bbox)
#delete unuseful index for unmatched_trackers
for delete_index in embedding_unmatched_trackers:
delete_id = track_feature_list[delete_index][2048]
del self.track_id_dict[delete_id]
#self.tracker.delete_id(delete_id)
bbox_list = []
for id, item in self.track_id_dict.items():
if item['exist'] == True:
bbox_list.append(item['bbox_and_score'] + [id])
return bbox_list
def match_detection_tracking_iou_embedding(self, detections, track_list,
frame):
rgb_frame = frame[:, :, ::-1]
#print(detections)
for track in track_list:
track_score = track[4]
if track_score >= self.effective_track_thresh:
detections.append(track[0:5])
#print(detections)
detections, keypoint_list = self.oks_filter(detections, frame)
#get feature for former trackers
database_bbox_list = []
for database_id in self.track_id_dict:
database_id_bbox = self.track_id_dict[database_id][
'bbox_and_score'] + [database_id]
database_bbox_list.append(database_id_bbox)
matches, unmatched_detections, unmatched_trackers = self.matcher.associate_detections_to_trackers_iou(
detections,
database_bbox_list,
iou_threshold=self.iou_match_thresh)
#update the matched trackers with detection bbox
for match in matches:
det_index, track_index = match
det_bbox = detections[det_index]
update_id = database_bbox_list[track_index][5]
self.update_id(frame, rgb_frame, det_bbox, update_id)
#final_bbox.append(det_bbox+[update_id])
#create new index for unmatched_detections
det_feature_list = []
for new_index in unmatched_detections:
det_bbox = detections[new_index]
det_score = det_bbox[4]
det_id_feature = self.embedding(frame, det_bbox) + det_bbox
det_feature_list.append(det_id_feature)
track_feature_list = []
for delete_index in unmatched_trackers:
track_bbox = database_bbox_list[delete_index]
track_score, delete_id = track_bbox[4], track_bbox[5]
delete_id_feature = self.track_id_dict[delete_id]['feature'] + [
delete_id
]
track_feature_list.append(delete_id_feature)
embedding_matches, \
embedding_unmatched_detections,\
embedding_unmatched_trackers = self.matcher.associate_detections_to_trackers_embedding(det_feature_list,
track_feature_list,
distance_threshold = self.embedding_match_thresh)
#update matched embedding detections and former tracking feature
for match in embedding_matches:
det_index, track_index = match
det_bbox = det_feature_list[det_index][2048:]
update_id = track_feature_list[track_index][2048]
self.update_id(frame, rgb_frame, det_bbox, update_id)
#create new id for unmatched detections
for new_index in embedding_unmatched_detections:
det_bbox = det_feature_list[new_index][2048:]
det_score = det_bbox[4]
pose_position, pose_value, pose_heatmap = self.pose_detect(
frame, det_bbox)
if det_score >= self.effective_detection_thresh and np.sum(
pose_value >= self.effective_keypoints_thresh
) >= self.effective_keypoints_number:
self.create_id(frame, rgb_frame, det_bbox)
#delete unuseful index for unmatched_trackers
for delete_index in embedding_unmatched_trackers:
delete_id = track_feature_list[delete_index][2048]
del self.track_id_dict[delete_id]
self.tracker.delete_id(delete_id)
bbox_list = []
for id, item in self.track_id_dict.items():
if item['exist'] == True:
bbox_list.append(item['bbox_and_score'] + [id])
return bbox_list
def match_and_track_embedding_no_database(self, detections, track_list,
frame):
# print(self.track_id_dict.keys())
rgb_frame = frame[:, :, ::-1]
matches, unmatched_detections, unmatched_trackers = self.matcher.associate_detections_to_trackers_iou(
detections, track_list, iou_threshold=self.iou_match_thresh)
has_tracked_id = set()
for match in matches:
det_index, track_index = match
has_tracked_id.add(track_index)
det_bbox = detections[det_index]
update_id = track_list[track_index][5]
self.update_id(frame, rgb_frame, det_bbox, update_id)
#get feature unmatched_detections
det_feature_list = []
for new_index in unmatched_detections:
det_bbox = detections[new_index]
det_score = det_bbox[4]
det_id_feature = self.embedding(frame, det_bbox) + det_bbox
det_feature_list.append(det_id_feature)
#get feature for unmatched_trackers
track_feature_list = []
for delete_index in unmatched_trackers:
track_bbox = track_list[delete_index]
track_score, delete_id = track_bbox[4], track_bbox[5]
delete_id_feature = self.track_id_dict[delete_id]['feature'] + [
delete_id
]
track_feature_list.append(delete_id_feature)
#match the detection and tracklist
embedding_matches, \
embedding_unmatched_detections,\
embedding_unmatched_trackers = self.matcher.associate_detections_to_trackers_embedding(det_feature_list,
track_feature_list,
distance_threshold = self.embedding_match_thresh)
#update matched embedding detections and former tracking feature
for match in embedding_matches:
det_index, track_index = match
det_bbox = det_feature_list[det_index][2048:]
update_id = track_feature_list[track_index][2048]
self.update_id(frame, rgb_frame, det_bbox, update_id)
#create new id for unmatched detections
for new_index in embedding_unmatched_detections:
det_bbox = det_feature_list[new_index][2048:]
det_score = det_bbox[4]
pose_position, pose_value, pose_heatmap = self.pose_detect(
frame, det_bbox)
if det_score >= self.effective_detection_thresh and np.sum(
pose_value >= self.effective_keypoints_thresh
) >= self.effective_keypoints_number:
self.create_id(frame, rgb_frame, det_bbox)
#delete unuseful index for unmatched_trackers
for delete_index in embedding_unmatched_trackers:
delete_id = track_feature_list[delete_index][2048]
del self.track_id_dict[delete_id]
self.tracker.delete_id(delete_id)
bbox_list = []
for id, item in self.track_id_dict.items():
if item['exist'] == True:
bbox_list.append(item['bbox_and_score'] + [id])
#print(bbox_list)
return bbox_list
def match_and_track_embedding_temporal_database(self, detections,
track_list, frame):
# print(self.track_id_dict.keys())
rgb_frame = frame[:, :, ::-1]
matches, unmatched_detections, unmatched_trackers = self.matcher.associate_detections_to_trackers_iou(
detections, track_list, iou_threshold=self.iou_match_thresh)
has_tracked_id = set()
for match in matches:
det_index, track_index = match
has_tracked_id.add(track_index)
det_bbox = detections[det_index]
update_id = track_list[track_index][5]
self.update_id(frame, rgb_frame, det_bbox, update_id)
#get feature unmatched_detections
det_feature_list = []
for new_index in unmatched_detections:
det_bbox = detections[new_index]
det_score = det_bbox[4]
det_id_feature = self.embedding(frame, det_bbox) + det_bbox
det_feature_list.append(det_id_feature)
#get feature for unmatched_trackers
track_feature_list = []
for delete_index in unmatched_trackers:
track_bbox = track_list[delete_index]
track_score, delete_id = track_bbox[4], track_bbox[5]
delete_id_features = self.track_id_dict[delete_id]['feature']
for delete_id_feature in delete_id_features:
track_feature_list.append(delete_id_feature + [delete_id])
#match the detection and tracklist
embedding_matches, \
embedding_unmatched_detections,\
embedding_unmatched_trackers = self.matcher.associate_detections_to_trackers_embedding(det_feature_list,
track_feature_list,
distance_threshold = self.embedding_match_thresh)
#update matched embedding detections and former tracking feature
update_ids = dict()
for match in embedding_matches:
det_index, track_index = match
update_id = track_feature_list[track_index][2048]
if not update_id in update_ids:
update_ids[update_id] = det_index
else:
score_former = det_feature_list[update_ids[update_id]][-1]
score_now = det_feature_list[det_index][-1]
if score_now >= score_former:
update_ids[update_id] = det_index
for update_id, det_index in update_ids.items():
det_bbox = det_feature_list[det_index][2048:]
self.update_id(frame, rgb_frame, det_bbox, update_id)
#create new id for unmatched detections
for new_index in embedding_unmatched_detections:
det_bbox = det_feature_list[new_index][2048:]
det_score = det_bbox[4]
pose_position, pose_value, pose_heatmap = self.pose_detect(
frame, det_bbox)
if det_score >= self.effective_detection_thresh and np.sum(
pose_value >= self.effective_keypoints_thresh
) >= self.effective_keypoints_number:
self.create_id(frame, rgb_frame, det_bbox)
#delete unuseful index for unmatched_trackers
for delete_index in embedding_unmatched_trackers:
delete_id = track_feature_list[delete_index][2048]
if delete_id not in update_ids and delete_id in self.track_id_dict:
del self.track_id_dict[delete_id]
self.tracker.delete_id(delete_id)
bbox_list = []
for id, item in self.track_id_dict.items():
if item['exist'] == True:
bbox_list.append(item['bbox_and_score'] + [id])
#print(bbox_list)
return bbox_list
def match_and_track_embedding_no_database_nms(self, detections, track_list,
frame):
# print(self.track_id_dict.keys())
rgb_frame = frame[:, :, ::-1]
all_bbox = []
for det in detections:
all_bbox.append(det)
for track in track_list:
if not track[4] >= self.effective_track_thresh:
continue
all_bbox.append(track[0:4] + [track[5]])
all_bbox = np.array(all_bbox)
#print(all_bbox)
keep = self.matcher.nms(all_bbox, self.nms_thresh)
keep_bboxes = all_bbox[keep].tolist()
#print(len(all_bbox), len(keep_bboxes))
#print(keep_bboxes)
#get feature unmatched_detections
det_feature_list = []
for keep_bbox in keep_bboxes:
det_id_feature = self.embedding(frame, keep_bbox) + keep_bbox
det_feature_list.append(det_id_feature)
#get feature for unmatched_trackers
track_feature_list = []
for former_id in self.track_id_dict:
former_id_feature = self.track_id_dict[former_id]['feature'] + [
former_id
]
track_feature_list.append(former_id_feature)
#match the detection and tracklist
embedding_matches, \
embedding_unmatched_detections,\
embedding_unmatched_trackers = self.matcher.associate_detections_to_trackers_embedding(det_feature_list,
track_feature_list,
distance_threshold = self.embedding_match_thresh)
#update matched embedding detections and former tracking feature
for match in embedding_matches:
det_index, track_index = match
det_bbox = det_feature_list[det_index][2048:]
update_id = track_feature_list[track_index][2048]
self.update_id(frame, rgb_frame, det_bbox, update_id)
#create new id for unmatched detections
for new_index in embedding_unmatched_detections:
det_bbox = det_feature_list[new_index][2048:]
det_score = det_bbox[4]
pose_position, pose_value, pose_heatmap = self.pose_detect(
frame, det_bbox)
if det_score >= self.effective_detection_thresh and np.sum(
pose_value >= self.effective_keypoints_thresh
) >= self.effective_keypoints_number:
self.create_id(frame, rgb_frame, det_bbox)
#delete unuseful index for unmatched_trackers
for delete_index in embedding_unmatched_trackers:
delete_id = track_feature_list[delete_index][2048]
del self.track_id_dict[delete_id]
self.tracker.delete_id(delete_id)
bbox_list = []
for id, item in self.track_id_dict.items():
if item['exist'] == True:
bbox_list.append(item['bbox_and_score'] + [id])
#print(bbox_list)
return bbox_list
def match_and_track_iou(self, detections, track_list, frame):
rgb_frame = frame[:, :, ::-1]
matches, unmatched_detections, unmatched_trackers = self.matcher.associate_detections_to_trackers_iou(
detections, track_list, iou_threshold=self.iou_match_thresh)
#update the matched trackers with detection bbox
for match in matches:
det_index, track_index = match
det_bbox = detections[det_index]
update_id = track_list[track_index][5]
#print('id {} has been updated'.format(update_id))
self.track_id_dict[update_id]['bbox_and_score'] = det_bbox[0:5]
self.update_tracker(rgb_frame, det_bbox, update_id)
#create new index for unmatched_detections
for new_index in unmatched_detections:
det_bbox = detections[new_index]
det_score = det_bbox[4]
pose_position, pose_value, pose_heatmap = self.pose_detect(
frame, det_bbox)
if det_score >= self.effective_detection_thresh and np.sum(
pose_value >= self.effective_keypoints_thresh
) >= self.effective_keypoints_number:
#print('add id{} now'.format(self.new_id_flag))
self.create_id(frame, rgb_frame, det_bbox)
#delete unuseful index for unmatched_trackers
for delete_index in unmatched_trackers:
track_bbox = track_list[delete_index]
track_score, delete_id = track_bbox[4], track_bbox[5]
#pose_position, pose_value, pose_heatmap = self.pose_detect(frame, track_bbox)
# if track_score >= self.effective_track_thresh and np.sum(pose_value >= self.effective_keypoints_thresh) >= self.effective_keypoints_number:
# self.track_id_dict[delete_id]['bbox_and_score']=track_bbox[0:5]
# else:
# #print('delete id{} now'.format(delete_id))
del self.track_id_dict[delete_id]
self.tracker.delete_id(delete_id)
bbox_list = []
for id, item in self.track_id_dict.items():
bbox_list.append(item['bbox_and_score'] + [id])
#print(bbox_list)
return bbox_list
#bbox must be format of x1y1x2y2
def update_tracker(self, rgb_frame, bbox, track_id):
self.tracker.update_data_dict(rgb_frame, bbox, track_id)
def embedding(self, frame, bbox):
start = time.time()
#print(bbox)
feature = self.embedder.get_feature(frame, bbox)
#print('Embedding takes {}s'.format(time.time()-start))
return feature
def detect(self, im):
return self.detector.detect(im, self.nms_thresh, 0.5)
def pose_detect(self, im, bbox):
return self.posenet.detect_pose(im, bbox)
class Track_And_Detect(object):
effective_track_thresh = 0.5
effective_detection_thresh = 0.5
effective_keypoints_thresh = 0.6
effective_keypoints_number = 8
iou_match_thresh = 0.5
nms_thresh = 0.5
oks_thresh = 0.8
embedding_match_thresh = 2
feature_length = 2048
oks_flag = True
tracker_flag = True
tracker_update_flag = True
new_embedding_flag = True
descrease_tracker_flag = True
#descrease_tracker_flag = False
def __init__(self,
gpu_id=0,
track_model=None,
pose_model=None,
embedding_model=None):
if self.tracker_flag:
self.tracker = SiamFCTracker(gpu_id, track_model)
self.posenet = PoseNet(gpu_id, pose_model)
self.matcher = Matcher()
print('----------------------------------------')
print('Flag parameters are set as follow:')
print('Tracker flag: {}'.format(self.tracker_flag))
print('Tracker update flag: {}'.format(self.tracker_update_flag))
print('Decrease tracker flag: {}'.format(self.descrease_tracker_flag))
print('New embedding(with pose) flag: {}'.format(
self.new_embedding_flag))
print('----------------------------------------')
# bbox must be format of x1y1x2y2
def update_tracker(self, rgb_frame, bbox, track_id):
self.tracker.update_data_dict(rgb_frame, bbox, track_id)
def pose_detect(self, im, bbox):
return self.posenet.detect_pose(im, bbox)
def embedding(self, frame, bbox):
feature = self.posenet.embedding(frame, bbox)
return feature
# initialize the first frame of this video
def init_tracker(self, frame, bbox_list):
self.new_id_flag = 0
self.track_id_dict = dict()
if self.tracker_flag:
self.tracker.clear_data()
# conver bgr(opencv) to rgb
rgb_frame = frame[:, :, ::-1]
bbox_list, keypoint_list = self.oks_filter(bbox_list, frame)
for bbox in bbox_list:
self.create_id(frame, rgb_frame, bbox)
bbox_list = []
for id, item in self.track_id_dict.items():
bbox = item['bbox_and_score'] + [id]
bbox_list.append(bbox)
return bbox_list
def oks_filter(self, det_list, frame):
keypoint_list = []
for bbox in det_list:
center, scale = self.posenet.x1y1x2y2_to_cs(bbox[0:4])
area = np.prod(scale * 200, 1)
pred = np.zeros((15, 3), dtype=np.float32)
pose_positions, pose_vals, pose_heatmaps = self.pose_detect(
frame, bbox)
pred[:, 0:2] = pose_positions
pred[:, 2] = pose_vals
score_all, valid_num = 0, 0
for i in range(15):
score_i = pose_vals[i]
if score_i >= 0.2:
score_all += score_i
valid_num += 1
if valid_num != 0:
new_score = score_all / valid_num * bbox[4]
else:
new_score = 0
keypoint_dict = {
'score': new_score,
'area': area,
'keypoints': pred
}
keypoint_list.append(keypoint_dict)
keep = self.matcher.oks_nms(keypoint_list, thresh=self.oks_thresh)
new_det_list = [det_list[i] for i in keep]
new_keypoint_list = [keypoint_list[i] for i in keep]
return new_det_list, new_keypoint_list
def create_id(self, frame, rgb_frame, bbox):
score = bbox[4]
bbox = bbox[0:4]
track_id = self.new_id_flag
if self.new_embedding_flag:
feature = self.embedding(frame, bbox)
else:
feature = self.embedder.embedding(frame, bbox)
self.track_id_dict[track_id] = {
'bbox_and_score': bbox + [score],
'feature': feature,
'frame_flag': 1,
'exist': True
}
if self.tracker_flag:
self.update_tracker(rgb_frame, bbox, track_id)
self.new_id_flag += 1
def update_id(self, frame, rgb_frame, det_bbox, track_id):
bbox, score = det_bbox[0:4], det_bbox[4]
if self.new_embedding_flag:
feature = np.array(self.embedding(frame, bbox))
else:
feature = np.array(self.embedder.embedding(frame, bbox))
former_track_dict = self.track_id_dict[track_id]
former_frame_flag, former_feature = former_track_dict[
'frame_flag'], np.array(former_track_dict['feature'])
now_frame_flag = former_frame_flag + 1
now_feature = feature.tolist()
self.track_id_dict[track_id] = {
'bbox_and_score': det_bbox,
'feature': now_feature,
'frame_flag': now_frame_flag,
'exist': True
}
if self.tracker_flag and self.tracker_update_flag:
self.update_tracker(rgb_frame, bbox, track_id)
def multi_track(self, frame):
rgb_frame = frame[:, :, ::-1]
bbox_list = []
for id in self.track_id_dict:
if self.track_id_dict[id]['exist'] == False:
continue
bbox, score = self.tracker.track_id(rgb_frame, id)
bbox_list.append(bbox + [score] + [id])
self.track_id_dict[id]['bbox'] = bbox
return bbox_list
def match_detection_tracking_oks_iou_embedding(self, detections,
track_list, frame):
rgb_frame = frame[:, :, ::-1]
#print(detections)
for track in track_list:
track_score = track[4]
if track_score >= self.effective_track_thresh:
if self.descrease_tracker_flag:
track_score -= 0.35
detections.append(track[0:4] + [track_score])
#print(detections)
if self.oks_flag:
detections, keypoint_list = self.oks_filter(detections, frame)
#decrease the tracking score
#get feature for former trackers
database_bbox_list = []
for database_id in self.track_id_dict:
database_id_bbox = self.track_id_dict[database_id][
'bbox_and_score'] + [database_id]
database_bbox_list.append(database_id_bbox)
matches, unmatched_detections, unmatched_trackers = self.matcher.associate_detections_to_trackers_iou(
detections,
database_bbox_list,
iou_threshold=self.iou_match_thresh)
#update the matched trackers with detection bbox
for match in matches:
det_index, track_index = match
det_bbox = detections[det_index]
update_id = database_bbox_list[track_index][5]
self.update_id(frame, rgb_frame, det_bbox, update_id)
#create new index for unmatched_detections
det_feature_list = []
for new_index in unmatched_detections:
det_bbox = detections[new_index]
det_score = det_bbox[4]
if self.new_embedding_flag:
det_id_feature = self.embedding(frame, det_bbox) + det_bbox
else:
det_id_feature = self.embedder.embedding(frame,
det_bbox) + det_bbox
det_feature_list.append(det_id_feature)
track_feature_list = []
for delete_index in unmatched_trackers:
track_bbox = database_bbox_list[delete_index]
track_score, delete_id = track_bbox[4], track_bbox[5]
delete_id_feature = self.track_id_dict[delete_id]['feature'] + [
delete_id
]
track_feature_list.append(delete_id_feature)
embedding_matches, \
embedding_unmatched_detections,\
embedding_unmatched_trackers = self.matcher.associate_detections_to_trackers_embedding(det_feature_list,
track_feature_list,
distance_threshold = self.embedding_match_thresh)
#update matched embedding detections and former tracking feature
for match in embedding_matches:
det_index, track_index = match
det_bbox = det_feature_list[det_index][2048:]
update_id = track_feature_list[track_index][2048]
self.update_id(frame, rgb_frame, det_bbox, update_id)
#create new id for unmatched detections
for new_index in embedding_unmatched_detections:
det_bbox = det_feature_list[new_index][2048:]
det_score = det_bbox[4]
pose_position, pose_value, pose_heatmap = self.pose_detect(
frame, det_bbox)
if det_score >= self.effective_detection_thresh and np.sum(
pose_value >= self.effective_keypoints_thresh
) >= self.effective_keypoints_number:
self.create_id(frame, rgb_frame, det_bbox)
#delete unuseful index for unmatched_trackers
for delete_index in embedding_unmatched_trackers:
delete_id = track_feature_list[delete_index][2048]
del self.track_id_dict[delete_id]
if self.tracker_flag:
self.tracker.delete_id(delete_id)
bbox_list = []
for id, item in self.track_id_dict.items():
if item['exist'] == True:
bbox_list.append(item['bbox_and_score'] + [id])
return bbox_list
MR = dict(
(m, prefs.split(', ')) for [m, prefs] in (line.rstrip().split(': ')
for line in open('mentors.txt')))
# the women and their list of ordered spousal preferences
ME = dict(
(m, prefs.split(', ')) for [m, prefs] in (line.rstrip().split(': ')
for line in open('mentees.txt')))
previous_mentee_size = None
final_mentee_mentors = dict()
i = 0
while len(ME) > 0:
i += 1
previous_mentee_size = len(ME)
match = Matcher(MR, ME)
matches = match()
final_mentee_mentors.update(matches)
for mentee, mentor in matches.items():
# This mentee already found a mentor for themselves. Delete from matching.
ME.pop(mentee, True)
# Delete this mentee from all mentor selections, so that mentors can fall back to their other
# selections.
# todo what to do with mentees that were not selected by any mentor?
new_MR = MR.copy()
for mrk, mrv in MR.items():
if mentee in mrv:
new_MR.get(mrk).remove(mentee)
from match import Matcher
# the men and their list of ordered spousal preferences
M = dict(
(m, prefs.split(', '))
for [m, prefs] in (line.rstrip().split(': ') for line in open('men.txt')))
# the women and their list of ordered spousal preferences
W = dict(
(m, prefs.split(', ')) for [m, prefs] in (line.rstrip().split(': ')
for line in open('women.txt')))
# initialize Matcher with preference lists for both men and women
match = Matcher(M, W)
# check if the mapping of wives to husbands is stable
def is_stable(wives, verbose=False):
for w, m in wives.items():
i = M[m].index(w)
preferred = M[m][:i]
for p in preferred:
h = wives[p]
if W[p].index(m) < W[p].index(h):
msg = "{}'s marriage to {} is unstable: " + \
"{} prefers {} over {} and {} prefers " + \
"{} over her current husband {}"
if verbose:
print msg.format(m, w, m, p, w, p, m, h)
return False
return True
import cv2
import imutils
import numpy as np
import rospy
from geometry_msgs.msg import Point
from std_msgs.msg import String
from background_subtraction import BackgroundSubtractor
from match import Matcher
MIN_AREA = 10000
MAX_DIST = 10
k_dilate = cv2.getStructuringElement(cv2.MORPH_DILATE, (5, 5), (2, 2))
matcher = Matcher()
def within(low, x, high):
return low < x and x < high
def x2t(x, dim, fov):
theta = fov / 2
# angle difference
a = float(x) / dim # assuming 640 is frame width
return np.arctan2(2 * a * np.sin(theta) - np.sin(theta), np.cos(theta))
def cnt_avg_col(cnt, img):
mask = np.zeros(img.shape[:-1], np.uint8)
for w, m in wives.items():
i = M[m].index(w)
preferred = M[m][:i]
for p in preferred:
h = wives[p]
if W[p].index(m) < W[p].index(h):
msg = "{}'s marriage to {} is unstable: " + \
"{} prefers {} over {} and {} prefers " + \
"{} over her current husband {}"
if verbose:
print msg.format(m, w, m, p, w, p, m, h)
return False
return True
# initialize Matcher with preference lists for both men and women
match = Matcher(M, W)
# match men and women; returns a mapping of wives to husbands
wives = match()
assert match.is_stable() # should be a stable matching
assert wives['w1'] is 'm3' # `w1` is married to `m3`
# now change prefs of woman `w1` and rematch
W['w1'] = ["m3", "m1", "m2"]
# re-initialize Matcher with revised preference lists
match = Matcher(M, W)
wives = match()
assert match.is_stable() # should be a stable matching
assert wives['w1'] is 'm1' # but `w1` is now married to `m1`
from match import Matcher
# the men and their list of ordered spousal preferences
M = dict((m, prefs.split(', ')) for [m, prefs] in (line.rstrip().split(': ')
for line in open('men.txt')))
# the women and their list of ordered spousal preferences
W = dict((m, prefs.split(', ')) for [m, prefs] in (line.rstrip().split(': ')
for line in open('women.txt')))
# for each man construct a random list of forbidden wives
forbidden = {} # { 'dan': ['gay', 'eve', 'abi'], 'hal': ['eve'] }
for m, prefs in M.items():
n = random.randint(0, len(prefs) - 1)
forbidden[m] = random.sample(prefs, n) # random sample of n wives
match = Matcher(M, W, forbidden)
# match men and women; returns a mapping of wives to husbands
wives = match()
assert match.is_stable(wives) # should be a stable matching
# swap the husbands of two wives, which should make the matching unstable
a, b = random.sample(wives.keys(), 2)
wives[b], wives[a] = wives[a], wives[b]
match.is_stable(wives, verbose=True)
