def convert_named_chips(db_dir, img_dpath=None):
print('\n --- Convert Named Chips ---')
# --- Initialize ---
gt_format = '{}_{:d}.jpg'
print('gt_format (name, num) = %r' % gt_format)
if img_dpath is None:
img_dpath = db_dir + '/images'
print('Converting db_dir=%r and img_dpath=%r' % (db_dir, img_dpath))
# --- Build Image Table ---
helpers.print_('Building name table: ')
gx2_gname = helpers.list_images(img_dpath)
gx2_gid = range(1, len(gx2_gname) + 1)
print('There are %d images' % len(gx2_gname))
# ---- Build Name Table ---
helpers.print_('Building name table: ')
name_set = set([])
for gx, gname in enumerate(gx2_gname):
name, num = parse.parse(gt_format, gname)
name_set.add(name)
nx2_name = ['____', '____'] + list(name_set)
nx2_nid = [1, 1] + range(2, len(name_set) + 2)
print('There are %d names' % (len(nx2_name) - 2))
# ---- Build Chip Table ---
print('[converdb] Building chip table: ')
cx2_cid = []
cx2_theta = []
cx2_roi = []
cx2_nx = []
cx2_gx = []
cid = 1
def add_to_hs_tables(gname, name, roi, theta=0):
cid = len(cx2_cid) + 1
nx = nx2_name.index(name)
gx = gx2_gname.index(gname)
cx2_cid.append(cid)
cx2_roi.append(roi)
cx2_nx.append(nx)
cx2_gx.append(gx)
cx2_theta.append(theta)
return cid
for gx, gname in enumerate(gx2_gname):
name, num = parse.parse(gt_format, gname)
img_fpath = join(img_dpath, gname)
(w, h) = Image.open(img_fpath).size
roi = [1, 1, w, h]
cid = add_to_hs_tables(gname, name, roi)
cx2_nid = np.array(nx2_nid)[cx2_nx]
cx2_gid = np.array(gx2_gid)[cx2_gx]
print('There are %d chips' % (cid - 1))
# Write tables
internal_dir = join(db_dir, ld2.RDIR_INTERNAL2)
helpers.ensurepath(internal_dir)
write_chip_table(internal_dir, cx2_cid, cx2_gid, cx2_nid, cx2_roi,
cx2_theta)
write_name_table(internal_dir, nx2_nid, nx2_name)
write_image_table(internal_dir, gx2_gid, gx2_gname)
def convert_named_chips(db_dir, img_dpath=None):
print('\n --- Convert Named Chips ---')
# --- Initialize ---
gt_format = '{}_{:d}.jpg'
print('gt_format (name, num) = %r' % gt_format)
if img_dpath is None:
img_dpath = db_dir + '/images'
print('Converting db_dir=%r and img_dpath=%r' % (db_dir, img_dpath))
# --- Build Image Table ---
helpers.print_('Building name table: ')
gx2_gname = helpers.list_images(img_dpath)
gx2_gid = range(1, len(gx2_gname) + 1)
print('There are %d images' % len(gx2_gname))
# ---- Build Name Table ---
helpers.print_('Building name table: ')
name_set = set([])
for gx, gname in enumerate(gx2_gname):
name, num = parse.parse(gt_format, gname)
name_set.add(name)
nx2_name = ['____', '____'] + list(name_set)
nx2_nid = [1, 1] + range(2, len(name_set) + 2)
print('There are %d names' % (len(nx2_name) - 2))
# ---- Build Chip Table ---
print('[converdb] Building chip table: ')
cx2_cid = []
cx2_theta = []
cx2_roi = []
cx2_nx = []
cx2_gx = []
cid = 1
def add_to_hs_tables(gname, name, roi, theta=0):
cid = len(cx2_cid) + 1
nx = nx2_name.index(name)
gx = gx2_gname.index(gname)
cx2_cid.append(cid)
cx2_roi.append(roi)
cx2_nx.append(nx)
cx2_gx.append(gx)
cx2_theta.append(theta)
return cid
for gx, gname in enumerate(gx2_gname):
name, num = parse.parse(gt_format, gname)
img_fpath = join(img_dpath, gname)
(w, h) = Image.open(img_fpath).size
roi = [1, 1, w, h]
cid = add_to_hs_tables(gname, name, roi)
cx2_nid = np.array(nx2_nid)[cx2_nx]
cx2_gid = np.array(gx2_gid)[cx2_gx]
print('There are %d chips' % (cid - 1))
# Write tables
internal_dir = join(db_dir, ld2.RDIR_INTERNAL2)
helpers.ensurepath(internal_dir)
write_chip_table(internal_dir, cx2_cid, cx2_gid, cx2_nid, cx2_roi, cx2_theta)
write_name_table(internal_dir, nx2_nid, nx2_name)
write_image_table(internal_dir, gx2_gid, gx2_gname)
def init_database_from_images(db_dir,
img_dpath=None,
gt_format=None,
allow_unknown_chips=False):
# --- Initialize ---
if img_dpath is None:
img_dpath = db_dir + '/images'
print('Converting db_dir=%r and img_dpath=%r' % (db_dir, img_dpath))
gx2_gid, gx2_gname = imagetables_from_img_dpath(img_dpath)
name_set = groundtruth_from_imagenames(gx2_gname, gt_format)
nx2_name, nx2_nid = nametables_from_nameset(name_set)
# ---- Build Chip Table ---
helpers.print_('Building chip table: ')
cx2_cid = []
cx2_theta = []
cx2_roi = []
cx2_nx = []
cx2_gx = []
cid = 1
def add_to_hs_tables(gname, name, roi, theta=0):
cid = len(cx2_cid) + 1
nx = nx2_name.index(name)
gx = gx2_gname.index(gname)
cx2_cid.append(cid)
cx2_roi.append(roi)
cx2_nx.append(nx)
cx2_gx.append(gx)
cx2_theta.append(theta)
return cid
for gx, gname in enumerate(gx2_gname):
if gt_format is None:
name = '____'
else:
name, num = parse.parse(gt_format, gname)
if name == '____' and not allow_unknown_chips:
continue
img_fpath = join(img_dpath, gname)
roi = roi_from_imgsize(img_fpath)
if not roi is None:
cid = add_to_hs_tables(gname, name, roi)
cx2_nid = np.array(nx2_nid)[cx2_nx]
cx2_gid = np.array(gx2_gid)[cx2_gx]
print('There are %d chips' % (cid - 1))
# Write tables
internal_dir = join(db_dir, ld2.RDIR_INTERNAL2)
helpers.ensurepath(internal_dir)
write_chip_table(internal_dir, cx2_cid, cx2_gid, cx2_nid, cx2_roi,
cx2_theta)
write_name_table(internal_dir, nx2_nid, nx2_name)
write_image_table(internal_dir, gx2_gid, gx2_gname)
return True
def init_database_from_images(db_dir, img_dpath=None, gt_format=None,
allow_unknown_chips=False):
# --- Initialize ---
if img_dpath is None:
img_dpath = db_dir + '/images'
print('Converting db_dir=%r and img_dpath=%r' % (db_dir, img_dpath))
gx2_gid, gx2_gname = imagetables_from_img_dpath(img_dpath)
name_set = groundtruth_from_imagenames(gx2_gname, gt_format)
nx2_name, nx2_nid = nametables_from_nameset(name_set)
# ---- Build Chip Table ---
helpers.print_('Building chip table: ')
cx2_cid = []
cx2_theta = []
cx2_roi = []
cx2_nx = []
cx2_gx = []
cid = 1
def add_to_hs_tables(gname, name, roi, theta=0):
cid = len(cx2_cid) + 1
nx = nx2_name.index(name)
gx = gx2_gname.index(gname)
cx2_cid.append(cid)
cx2_roi.append(roi)
cx2_nx.append(nx)
cx2_gx.append(gx)
cx2_theta.append(theta)
return cid
for gx, gname in enumerate(gx2_gname):
if gt_format is None:
name = '____'
else:
name, num = parse.parse(gt_format, gname)
if name == '____' and not allow_unknown_chips:
continue
img_fpath = join(img_dpath, gname)
roi = roi_from_imgsize(img_fpath)
if not roi is None:
cid = add_to_hs_tables(gname, name, roi)
cx2_nid = np.array(nx2_nid)[cx2_nx]
cx2_gid = np.array(gx2_gid)[cx2_gx]
print('There are %d chips' % (cid - 1))
# Write tables
internal_dir = join(db_dir, ld2.RDIR_INTERNAL2)
helpers.ensurepath(internal_dir)
write_chip_table(internal_dir, cx2_cid, cx2_gid, cx2_nid, cx2_roi, cx2_theta)
write_name_table(internal_dir, nx2_nid, nx2_name)
write_image_table(internal_dir, gx2_gid, gx2_gname)
return True
def measure_feat_pairs(allres, orgtype='top_true'):
print('Measure ' + orgtype + ' pairs')
orgres = allres.__dict__[orgtype]
entropy_list = []
scale_list = []
score_list = []
lbl = 'Measuring ' + orgtype + ' pair '
fmt_str = helpers.make_progress_fmt_str(len(orgres), lbl)
rank_skips = []
gt_skips = []
for ix, (qcx, cx, score, rank) in enumerate(orgres.iter()):
helpers.print_(fmt_str % (ix + 1,))
# Skip low ranks
if rank > 5:
rank_skips.append(qcx)
continue
other_cxs = hs.get_other_indexed_cxs(qcx)
# Skip no groundtruth
if len(other_cxs) == 0:
gt_skips.append(qcx)
continue
res = qcx2_res[qcx]
# Get matching feature indexes
fm = res.cx2_fm[cx]
# Get their scores
fs = res.cx2_fs[cx]
# Get matching descriptors
printDBG('\nfm.shape=%r' % (fm.shape,))
desc1 = cx2_desc[qcx][fm[:, 0]]
desc2 = cx2_desc[cx][fm[:, 1]]
# Get matching keypoints
kpts1 = cx2_kpts[qcx][fm[:, 0]]
kpts2 = cx2_kpts[cx][fm[:, 1]]
# Get their scale
scale1_m = sv2.keypoint_scale(kpts1)
scale2_m = sv2.keypoint_scale(kpts2)
# Get their entropy
entropy1 = descriptor_entropy(desc1, bw_factor=1)
entropy2 = descriptor_entropy(desc2, bw_factor=1)
# Append to results
entropy_tup = np.array(zip(entropy1, entropy2))
scale_tup = np.array(zip(scale1_m, scale2_m))
entropy_tup = entropy_tup.reshape(len(entropy_tup), 2)
scale_tup = scale_tup.reshape(len(scale_tup), 2)
entropy_list.append(entropy_tup)
scale_list.append(scale_tup)
score_list.append(fs)
print('Skipped %d total.' % (len(rank_skips) + len(gt_skips),))
print('Skipped %d for rank > 5, %d for no gt' % (len(rank_skips), len(gt_skips),))
print(np.unique(map(len, entropy_list)))
def evstack(tup):
return np.vstack(tup) if len(tup) > 0 else np.empty((0, 2))
def ehstack(tup):
return np.hstack(tup) if len(tup) > 0 else np.empty((0, 2))
entropy_pairs = evstack(entropy_list)
scale_pairs = evstack(scale_list)
scores = ehstack(score_list)
print('\n * Measured %d pairs' % len(entropy_pairs))
return entropy_pairs, scale_pairs, scores
def __akmeans_iterate(data, clusters, datax2_clusterx_old, max_iters,
flann_params, ave_unchanged_thresh,
ave_unchanged_iterwin):
num_data = data.shape[0]
num_clusters = clusters.shape[0]
# Keep track of how many points have changed in each iteration
xx2_unchanged = np.zeros(ave_unchanged_iterwin, dtype=np.int32) + len(data)
print('[algos] Running akmeans: data.shape=%r ; num_clusters=%r' %
(data.shape, num_clusters))
print('[algos] * max_iters = %r ' % max_iters)
#print(' * dtype = %r ' % params.__BOW_DTYPE__)
print('[algos] * ave_unchanged_iterwin=%r ; ave_unchanged_thresh=%r' %
(ave_unchanged_thresh, ave_unchanged_iterwin))
print('[algos] Printing akmeans info in format:' +
'time (iterx, ave(#changed), #unchanged)')
for xx in xrange(0, max_iters):
# 1) Find each datapoints nearest cluster center
tt = helpers.tic()
helpers.print_('...tic')
helpers.flush()
(datax2_clusterx, _dist) = ann_flann_once(clusters, data, 1,
flann_params)
ellapsed = helpers.toc(tt)
helpers.print_('...toc(%.2fs)' % ellapsed)
helpers.flush()
# 2) Find new cluster datapoints
datax_sort = datax2_clusterx.argsort() # NOQA
clusterx_sort = datax2_clusterx[datax_sort]
_L = 0
clusterx2_dataLRx = [None for _ in xrange(num_clusters)]
for _R in xrange(len(datax_sort) + 1): # Slide R
if _R == num_data or clusterx_sort[_L] != clusterx_sort[_R]:
clusterx2_dataLRx[clusterx_sort[_L]] = (_L, _R)
_L = _R
# 3) Compute new cluster centers
helpers.print_('+')
helpers.flush()
for clusterx, dataLRx in enumerate(clusterx2_dataLRx):
if dataLRx is None:
continue # ON EMPTY CLUSTER
(_L, _R) = dataLRx
clusters[clusterx] = np.mean(data[datax_sort[_L:_R]], axis=0)
#if params.__BOW_DTYPE__ == np.uint8:
#clusters[clusterx] = np.array(np.round(clusters[clusterx]),
# dtype=params.__BOW_DTYPE__)
clusters[clusterx] = np.array(np.round(clusters[clusterx]),
dtype=np.uint8)
# 4) Check for convergence (no change of cluster id)
helpers.print_('+')
helpers.flush()
num_changed = (datax2_clusterx_old != datax2_clusterx).sum()
xx2_unchanged[xx % ave_unchanged_iterwin] = num_changed
ave_unchanged = xx2_unchanged.mean()
#(iterx, ave(#changed), #unchanged)
helpers.print_(' (%d, %.2f, %d)\n' % (xx, ave_unchanged, num_changed))
helpers.flush()
if ave_unchanged < ave_unchanged_thresh:
break
else: # Iterate
datax2_clusterx_old = datax2_clusterx
if xx % 5 == 0:
sys.stdout.flush()
print('[algos] * AKMEANS: converged in %d/%d iters' % (xx + 1, max_iters))
sys.stdout.flush()
return (datax2_clusterx, clusters)
def measure_feat_pairs(allres, orgtype='top_true'):
print('Measure ' + orgtype + ' pairs')
orgres = allres.__dict__[orgtype]
entropy_list = []
scale_list = []
score_list = []
lbl = 'Measuring ' + orgtype + ' pair '
fmt_str = helpers.make_progress_fmt_str(len(orgres), lbl)
rank_skips = []
gt_skips = []
for ix, (qcx, cx, score, rank) in enumerate(orgres.iter()):
helpers.print_(fmt_str % (ix + 1, ))
# Skip low ranks
if rank > 5:
rank_skips.append(qcx)
continue
other_cxs = hs.get_other_indexed_cxs(qcx)
# Skip no groundtruth
if len(other_cxs) == 0:
gt_skips.append(qcx)
continue
res = qcx2_res[qcx]
# Get matching feature indexes
fm = res.cx2_fm[cx]
# Get their scores
fs = res.cx2_fs[cx]
# Get matching descriptors
printDBG('\nfm.shape=%r' % (fm.shape, ))
desc1 = cx2_desc[qcx][fm[:, 0]]
desc2 = cx2_desc[cx][fm[:, 1]]
# Get matching keypoints
kpts1 = cx2_kpts[qcx][fm[:, 0]]
kpts2 = cx2_kpts[cx][fm[:, 1]]
# Get their scale
scale1_m = sv2.keypoint_scale(kpts1)
scale2_m = sv2.keypoint_scale(kpts2)
# Get their entropy
entropy1 = descriptor_entropy(desc1, bw_factor=1)
entropy2 = descriptor_entropy(desc2, bw_factor=1)
# Append to results
entropy_tup = np.array(zip(entropy1, entropy2))
scale_tup = np.array(zip(scale1_m, scale2_m))
entropy_tup = entropy_tup.reshape(len(entropy_tup), 2)
scale_tup = scale_tup.reshape(len(scale_tup), 2)
entropy_list.append(entropy_tup)
scale_list.append(scale_tup)
score_list.append(fs)
print('Skipped %d total.' % (len(rank_skips) + len(gt_skips), ))
print('Skipped %d for rank > 5, %d for no gt' % (
len(rank_skips),
len(gt_skips),
))
print(np.unique(map(len, entropy_list)))
def evstack(tup):
return np.vstack(tup) if len(tup) > 0 else np.empty((0, 2))
def ehstack(tup):
return np.hstack(tup) if len(tup) > 0 else np.empty((0, 2))
entropy_pairs = evstack(entropy_list)
scale_pairs = evstack(scale_list)
scores = ehstack(score_list)
print('\n * Measured %d pairs' % len(entropy_pairs))
return entropy_pairs, scale_pairs, scores
def __akmeans_iterate(data,
clusters,
datax2_clusterx_old,
max_iters,
flann_params,
ave_unchanged_thresh,
ave_unchanged_iterwin):
num_data = data.shape[0]
num_clusters = clusters.shape[0]
# Keep track of how many points have changed in each iteration
xx2_unchanged = np.zeros(ave_unchanged_iterwin, dtype=np.int32) + len(data)
print('[algos] Running akmeans: data.shape=%r ; num_clusters=%r' %
(data.shape, num_clusters))
print('[algos] * max_iters = %r ' % max_iters)
#print(' * dtype = %r ' % params.__BOW_DTYPE__)
print('[algos] * ave_unchanged_iterwin=%r ; ave_unchanged_thresh=%r' %
(ave_unchanged_thresh, ave_unchanged_iterwin))
print('[algos] Printing akmeans info in format:' +
'time (iterx, ave(#changed), #unchanged)')
for xx in xrange(0, max_iters):
# 1) Find each datapoints nearest cluster center
tt = helpers.tic()
helpers.print_('...tic')
helpers.flush()
(datax2_clusterx, _dist) = ann_flann_once(clusters, data, 1,
flann_params)
ellapsed = helpers.toc(tt)
helpers.print_('...toc(%.2fs)' % ellapsed)
helpers.flush()
# 2) Find new cluster datapoints
datax_sort = datax2_clusterx.argsort() # NOQA
clusterx_sort = datax2_clusterx[datax_sort]
_L = 0
clusterx2_dataLRx = [None for _ in xrange(num_clusters)]
for _R in xrange(len(datax_sort) + 1): # Slide R
if _R == num_data or clusterx_sort[_L] != clusterx_sort[_R]:
clusterx2_dataLRx[clusterx_sort[_L]] = (_L, _R)
_L = _R
# 3) Compute new cluster centers
helpers.print_('+')
helpers.flush()
for clusterx, dataLRx in enumerate(clusterx2_dataLRx):
if dataLRx is None:
continue # ON EMPTY CLUSTER
(_L, _R) = dataLRx
clusters[clusterx] = np.mean(data[datax_sort[_L:_R]], axis=0)
#if params.__BOW_DTYPE__ == np.uint8:
#clusters[clusterx] = np.array(np.round(clusters[clusterx]),
# dtype=params.__BOW_DTYPE__)
clusters[clusterx] = np.array(np.round(clusters[clusterx]),
dtype=np.uint8)
# 4) Check for convergence (no change of cluster id)
helpers.print_('+')
helpers.flush()
num_changed = (datax2_clusterx_old != datax2_clusterx).sum()
xx2_unchanged[xx % ave_unchanged_iterwin] = num_changed
ave_unchanged = xx2_unchanged.mean()
#(iterx, ave(#changed), #unchanged)
helpers.print_(' (%d, %.2f, %d)\n' % (xx, ave_unchanged, num_changed))
helpers.flush()
if ave_unchanged < ave_unchanged_thresh:
break
else: # Iterate
datax2_clusterx_old = datax2_clusterx
if xx % 5 == 0:
sys.stdout.flush()
print('[algos] * AKMEANS: converged in %d/%d iters' % (xx + 1, max_iters))
sys.stdout.flush()
return (datax2_clusterx, clusters)
