def rpn_compute_stats(queue=None, imdb_name=None, cfg=None, rpn_test_prototxt=None):
"""Compute mean stds for anchors
"""
cfg.TRAIN.HAS_RPN = True
cfg.TRAIN.BBOX_REG = False # applies only to R-FCN bbox regression
cfg.TRAIN.PROPOSAL_METHOD = 'gt'
cfg.TRAIN.IMS_PER_BATCH = 1
import caffe
_init_caffe(cfg)
# NOTE: the matlab implementation computes proposals on flipped images, too.
# We compute them on the image once and then flip the already computed
# proposals. This might cause a minor loss in mAP (less proposal jittering).
roidb, imdb = get_roidb(imdb_name)
print 'Loaded dataset `{:s}` for proposal generation'.format(imdb.name)
mean_file = os.path.join(imdb.cache_path, imdb.name + '_means.npy')
std_file = os.path.join(imdb.cache_path, imdb.name + '_stds.npy')
if os.path.exists(mean_file) and os.path.exists(std_file):
means = np.load(mean_file)
stds = np.load(std_file)
else:
# Load RPN and configure output directory
rpn_net = caffe.Net(rpn_test_prototxt, caffe.TEST)
# Generate proposals on the imdb
print 'start computing means/stds, it may take several minutes...'
if imdb_name.startswith('coco'):
means, stds = imdb_rpn_compute_stats(rpn_net, imdb, anchor_scales=(4, 8, 16, 32))
else:
means, stds = imdb_rpn_compute_stats(rpn_net, imdb, anchor_scales=(8, 16, 32))
np.save(mean_file, means)
np.save(std_file, stds)
queue.put({'means': means, 'stds': stds})
def rpn_compute_stats(queue=None, imdb_name=None, cfg=None, rpn_test_prototxt=None):
"""Compute mean stds for anchors
"""
cfg.TRAIN.HAS_RPN = True
cfg.TRAIN.BBOX_REG = False # applies only to R-FCN bbox regression
cfg.TRAIN.PROPOSAL_METHOD = 'gt'
cfg.TRAIN.IMS_PER_BATCH = 1
import caffe
_init_caffe(cfg)
# NOTE: the matlab implementation computes proposals on flipped images, too.
# We compute them on the image once and then flip the already computed
# proposals. This might cause a minor loss in mAP (less proposal jittering).
roidb, imdb = get_roidb(imdb_name)
print 'Loaded dataset `{:s}` for proposal generation'.format(imdb.name)
mean_file = os.path.join(imdb.cache_path, imdb.name + '_means.npy')
std_file = os.path.join(imdb.cache_path, imdb.name + '_stds.npy')
if os.path.exists(mean_file) and os.path.exists(std_file):
means = np.load(mean_file)
stds = np.load(std_file)
else:
# Load RPN and configure output directory
rpn_net = caffe.Net(rpn_test_prototxt, caffe.TEST)
# Generate proposals on the imdb
print 'start computing means/stds, it may take several minutes...'
if imdb_name.startswith('coco'):
means, stds = imdb_rpn_compute_stats(rpn_net, imdb, anchor_scales=(4, 8, 16, 32))
else:
means, stds = imdb_rpn_compute_stats(rpn_net, imdb, anchor_scales=(8, 16, 32))
np.save(mean_file, means)
np.save(std_file, stds)
queue.put({'means': means, 'stds': stds})