def _im_exfeat(net, im, roi, blob_names):
"""Detect object classes in an image given object proposals.
Arguments:
net (caffe.Net): Fast R-CNN network to use
im (ndarray): color image to test (in BGR order)
roi (ndarray): 1 x 4 array of the target roi
blob_names (list of str): list of feature blob names to be extracted
Returns:
features (dict of ndarray): {blob name: R x D array of features}
"""
im_blob, im_scales = get_image_blob(im)
blobs = {
'data': im_blob,
'im_info': np.array(
[[im_blob.shape[2], im_blob.shape[3], im_scales[0]]],
dtype=np.float32),
'rois': get_rois_blob(roi, im_scales),
}
# reshape network inputs
for k, v in blobs.iteritems():
net.blobs[k].reshape(*(v.shape))
# do forward
forward_kwargs = {k: v.astype(np.float32, copy=False)
for k, v in blobs.iteritems()}
net.forward(**forward_kwargs)
features = {blob: net.blobs[blob].data.copy() for blob in blob_names} \
if blob_names is not None else {}
return features
def _im_exfeat(net, im, roi, blob_names):
"""Detect object classes in an image given object proposals.
Arguments:
net (caffe.Net): Fast R-CNN network to use
im (ndarray): color image to test (in BGR order)
roi (ndarray): 1 x 4 array of the target roi
blob_names (list of str): list of feature blob names to be extracted
Returns:
features (dict of ndarray): {blob name: R x D array of features}
"""
im_blob, im_scales = get_image_blob(im)
blobs = {
'data':
im_blob,
'im_info':
np.array([[im_blob.shape[2], im_blob.shape[3], im_scales[0]]],
dtype=np.float32),
'rois':
get_rois_blob(roi, im_scales),
}
# reshape network inputs
for k, v in blobs.iteritems():
net.blobs[k].reshape(*(v.shape))
# do forward
forward_kwargs = {
k: v.astype(np.float32, copy=False)
for k, v in blobs.iteritems()
}
net.forward(**forward_kwargs)
features = {blob: net.blobs[blob].data.copy() for blob in blob_names} \
if blob_names is not None else {}
return features
def _im_detect(net, im, roidb, blob_names=None):
"""Detect object classes in an image given object proposals.
Arguments:
net (caffe.Net): Fast R-CNN network to use
im (ndarray): color image to test (in BGR order)
roidb (an roidb item): to provide gt_boxes if necessary
blob_names (list of str): list of feature blob names to be extracted
Returns:
boxes (ndarray): R x (4*K) array of predicted bounding boxes
scores (ndarray): R x K array of object class scores (K includes
background as object category 0)
features (dict of ndarray): {blob name: R x D array of features}
"""
im_blob, im_scales = get_image_blob(im)
assert len(im_scales) == 1, "Only single-image batch implemented"
blobs = {
'data': im_blob,
'im_info': np.array(
[[im_blob.shape[2], im_blob.shape[3], im_scales[0]]],
dtype=np.float32),
}
if 'gt_boxes' in net.blobs:
# Supply gt_boxes as input. Used to get pid_labels for proposals.
blobs['gt_boxes'] = get_gt_boxes_blob(
roidb['boxes'], roidb['gt_classes'], roidb['gt_pids'], im_scales)
# reshape network inputs
for k, v in blobs.iteritems():
net.blobs[k].reshape(*(v.shape))
# do forward
forward_kwargs = {k: v.astype(np.float32, copy=False)
for k, v in blobs.iteritems()}
blobs_out = net.forward(**forward_kwargs)
# unscale rois back to raw image space
rois = net.blobs['rois'].data.copy()
boxes = rois[:, 1:5] / im_scales[0]
if cfg.TEST.SVM:
# use the raw scores before softmax under the assumption they
# were trained as linear SVMs
scores = net.blobs['cls_score'].data
else:
# the first column of the pid_prob is the non-person box score
scores = blobs_out['pid_prob'][:, 0]
scores = scores[:, np.newaxis]
scores = np.hstack([scores, 1. - scores])
if cfg.TEST.BBOX_REG:
# Apply bounding-box regression deltas
box_deltas = blobs_out['bbox_pred']
# As we no longer scale and shift the bbox_pred weights when snapshot,
# we need to manually do this during test.
if cfg.TRAIN.BBOX_NORMALIZE_TARGETS and \
cfg.TRAIN.BBOX_NORMALIZE_TARGETS_PRECOMPUTED:
num_classes = box_deltas.shape[1] // 4
stds = np.tile(cfg.TRAIN.BBOX_NORMALIZE_STDS, num_classes)
means = np.tile(cfg.TRAIN.BBOX_NORMALIZE_MEANS, num_classes)
box_deltas = box_deltas * stds + means
boxes = bbox_transform_inv(boxes, box_deltas)
boxes = clip_boxes(boxes, im.shape)
else:
# Simply repeat the boxes, once for each class
boxes = np.tile(boxes, (1, scores.shape[1]))
features = {blob: net.blobs[blob].data.copy() for blob in blob_names} \
if blob_names is not None else {}
return boxes, scores, features
def _im_detect(net, im, roidb, blob_names=None):
"""Detect object classes in an image given object proposals.
Arguments:
net (caffe.Net): Fast R-CNN network to use
im (ndarray): color image to test (in BGR order)
roidb (an roidb item): to provide gt_boxes if necessary
blob_names (list of str): list of feature blob names to be extracted
Returns:
boxes (ndarray): R x (4*K) array of predicted bounding boxes
scores (ndarray): R x K array of object class scores (K includes
background as object category 0)
features (dict of ndarray): {blob name: R x D array of features}
"""
im_blob, im_scales = get_image_blob(im)
assert len(im_scales) == 1, "Only single-image batch implemented"
blobs = {
'data': im_blob,
'im_info': np.array(
[[im_blob.shape[2], im_blob.shape[3], im_scales[0]]],
dtype=np.float32),
}
if 'gt_boxes' in net.blobs:
# Supply gt_boxes as input. Used to get pid_labels for proposals.
blobs['gt_boxes'] = get_gt_boxes_blob(
roidb['boxes'], roidb['gt_classes'], roidb['gt_pids'], im_scales)
# reshape network inputs
for k, v in blobs.iteritems():
net.blobs[k].reshape(*(v.shape))
# do forward
forward_kwargs = {k: v.astype(np.float32, copy=False)
for k, v in blobs.iteritems()}
blobs_out = net.forward(**forward_kwargs)
# unscale rois back to raw image space
rois = net.blobs['rois'].data.copy()
boxes = rois[:, 1:5] / im_scales[0]
if cfg.TEST.SVM:
# use the raw scores before softmax under the assumption they
# were trained as linear SVMs
scores = net.blobs['cls_score'].data
else:
# the first column of the pid_prob is the non-person box score
scores = blobs_out['pid_prob'][:, 0]
scores = scores[:, np.newaxis]
scores = np.hstack([scores, 1. - scores])
if cfg.TEST.BBOX_REG:
# Apply bounding-box regression deltas
box_deltas = blobs_out['bbox_pred']
# As we no longer scale and shift the bbox_pred weights when snapshot,
# we need to manually do this during test.
if cfg.TRAIN.BBOX_NORMALIZE_TARGETS and \
cfg.TRAIN.BBOX_NORMALIZE_TARGETS_PRECOMPUTED:
num_classes = box_deltas.shape[1] // 4
stds = np.tile(cfg.TRAIN.BBOX_NORMALIZE_STDS, num_classes)
means = np.tile(cfg.TRAIN.BBOX_NORMALIZE_MEANS, num_classes)
box_deltas = box_deltas * stds + means
boxes = bbox_transform_inv(boxes, box_deltas)
boxes = clip_boxes(boxes, im.shape)
else:
# Simply repeat the boxes, once for each class
boxes = np.tile(boxes, (1, scores.shape[1]))
features = {blob: net.blobs[blob].data.copy() for blob in blob_names} \
if blob_names is not None else {}
return boxes, scores, features