def init_net(net1, net2):
# Warmup on a dummy image
im = 128 * np.ones((300, 500, 3), dtype=np.uint8)
if net1 is not None:
for i in xrange(2):
im_detect(net1, im)
print "\n\n***********************************\n\n"
if net2 is not None:
time.sleep(3)
for i in xrange(2):
im_detect(net2, im)
print "\n\nInit Net Done!\n\n"
time.sleep(3)
def _image2bbox_top1(net, im, classes, t_cls, NMS_THRESH=0.3, CONF_THRESH=0.8):
"""
Detect object classes in an image using pre-computed object proposals.
only highest-score one
(very suitable for model trained on cropped images by person detector)
(also for model trained on whole image but maybe not good for multi-people in the image)
"""
timer = Timer()
timer.tic()
scores, boxes = im_detect(net, im)
# ignore bg
bboxes = []
is_target = len(t_cls) > 0
for cls_ind, cls in enumerate(classes[1:]):
if is_target and cls not in t_cls:
continue
cls_ind += 1
cls_scores = scores[:, cls_ind]
order = cls_scores.argsort()[::-1]
obj_ind = order[0]
score = cls_scores[obj_ind]
bbox = boxes[obj_ind, 4 * cls_ind:4 * (cls_ind + 1)]
bbox = [int(b) for b in bbox]
bboxes.append((bbox, score, cls))
total_time = timer.toc(average=False)
print "Detection took %ss for %s object proposals" % (total_time,
boxes.shape[0])
return bboxes
def _demo4image(net,
im_path,
classes,
t_cls,
out_dire=None,
NMS_THRESH=0.3,
CONF_THRESH=0.8):
"""Detect object classes in an image using pre-computed object proposals."""
timer = Timer()
timer.tic()
print 'Demo for {}'.format(im_path)
im = cv2.imread(im_path)
scores, boxes = im_detect(net, im)
# ignore bg
is_target = len(t_cls) > 0
for cls_ind, cls in enumerate(classes[1:]):
if is_target and cls not in t_cls:
continue
cls_ind += 1
cls_boxes = boxes[:, 4 * cls_ind:4 * (cls_ind + 1)]
cls_scores = scores[:, cls_ind]
dets = np.hstack(
(cls_boxes, cls_scores[:, np.newaxis])).astype(np.float32)
keep = nms(dets, NMS_THRESH)
dets = dets[keep, :]
vis_detections(im, cls, dets, im_path, out_dire, thresh=CONF_THRESH)
total_time = timer.toc(average=False)
print "Detection took %ss for %s object proposals" % (total_time,
boxes.shape[0])
def _demo4image_2(net,
im_path,
classes,
t_cls,
out_dire=None,
NMS_THRESH=0.3,
CONF_THRESH=0.6,
im_ext=".jpg"):
"""Detect object classes in an image using pre-computed object proposals."""
timer = Timer()
timer.tic()
print 'Demo for {}'.format(im_path)
im = cv2.imread(im_path)
scores, boxes = im_detect(net, im)
# ignore bg
is_target = len(t_cls) > 0
for cls_ind, cls in enumerate(classes[1:]):
if is_target and cls not in t_cls:
continue
cls_ind += 1
cls_boxes = boxes[:, 4 * cls_ind:4 * (cls_ind + 1)]
cls_scores = scores[:, cls_ind]
dets = np.hstack(
(cls_boxes, cls_scores[:, np.newaxis])).astype(np.float32)
keep = nms(dets, NMS_THRESH)
dets = dets[keep, :]
inds = np.where(dets[:, -1] >= CONF_THRESH)[0]
if len(inds) == 0:
continue
for i in inds:
bbox = dets[i, :4]
bbox = [int(b) for b in bbox]
score = dets[i, -1]
p1 = (bbox[0], bbox[1])
p2 = (bbox[2], bbox[3])
cv2.rectangle(im, p1, p2, (38, 231, 16), 2)
p3 = (bbox[0], bbox[1] - 5)
cv2.putText(im, '{:s} {:s}'.format(cls, str(score),), p3, \
cv2.FONT_HERSHEY_SIMPLEX, .56, (123, 19, 208), 1)
im_name = im_path.rsplit("/", 1)[1]
im_name = im_name.rsplit(".", 1)[0]
if out_dire and len(out_dire) > 0:
im_path2 = out_dire + im_name + im_ext
cv2.imwrite(im_path2, im)
else:
cv2.imshow(im_name, im)
cv2.waitKey(1)
cv2.destroyAllWindows()
total_time = timer.toc(average=False)
print "Detection took %ss for %s object proposals" % (total_time,
boxes.shape[0])
def _demo4image2file(net,
im_path,
fhd,
classes,
t_cls,
NMS_THRESH=0.3,
CONF_THRESH=0.8):
"""
Detect object classes in an image using pre-computed object proposals.
And write the results of bboxes into file by file handler `fhd` object.
Format:
im_path [[score bbox cls] ...]
"""
timer = Timer()
timer.tic()
print 'Demo for {}'.format(im_path)
im = cv2.imread(im_path)
scores, boxes = im_detect(net, im)
# ignore bg
h, w, _ = im.shape
info = im_path.strip()
is_target = len(t_cls) > 0
for cls_ind, cls in enumerate(classes[1:]):
if is_target and cls not in t_cls:
continue
cls_ind += 1
cls_boxes = boxes[:, 4 * cls_ind:4 * (cls_ind + 1)]
cls_scores = scores[:, cls_ind]
dets = np.hstack(
(cls_boxes, cls_scores[:, np.newaxis])).astype(np.float32)
keep = nms(dets, NMS_THRESH)
dets = dets[keep, :]
inds = np.where(dets[:, -1] >= CONF_THRESH)[0]
if len(inds) == 0:
return
for i in inds:
score = dets[i, -1]
score = str(score)
bbox = dets[i, :4]
bbox = [int(b) for b in bbox]
bbox = [x1, y1, x2, y2]
bbox = [str(b) for b in bbox]
bbox = " ".join(bbox).strip()
info = info + " " + score + " " + bbox + " " + cls
fhd.write(info.strip() + "\n")
total_time = timer.toc(average=False)
print "Detection took %ss for %s object proposals" % (total_time,
boxes.shape[0])
def _image2bbox2(net, im, classes, target_cls, cls_type=True):
"""
Detect object classes in an image using pre-computed object proposals.
>= threshold (maybe empty)
"""
if cls_type:
nms_thres = cfg.TEST.P_NMS_THRES # person
conf_thres = cfg.TEST.P_CONF_THRES
else:
nms_thres = cfg.TEST.T_NMS_THRES # torso
conf_thres = cfg.TEST.T_CONF_THRES
timer = Timer()
timer.tic()
# forward
scores, boxes, = im_detect(net, im)
# nms and threshold
bboxes = []
has_target = len(target_cls) > 0
for ind, cls in enumerate(classes[1:]):
if has_target and cls not in target_cls:
continue
ind += 1 # ignore bg
boxes2 = boxes[:, 4 * ind:4 * (ind + 1)]
scores2 = scores[:, ind]
dets = np.hstack((boxes2, scores2[:, np.newaxis])).astype(np.float32)
keep = nms(dets, nms_thres)
dets = dets[keep, :]
inds = np.where(dets[:, -1] >= conf_thres)[0]
if len(inds) == 0:
break
for i in inds:
bbox = dets[i, :4]
bbox = [int(b) for b in bbox]
score = dets[i, -1]
bboxes.append((bbox, score, cls))
total_time = timer.toc(average=False)
print "Detection took %ss for %s object proposals" % (total_time,
boxes.shape[0])
if len(bboxes) <= 0:
return None
return bboxes
def _demo4image_top1(net,
im_path,
classes,
t_cls,
out_dire=None,
im_ext=".jpg"):
"""Detect object classes in an image using pre-computed object proposals."""
timer = Timer()
timer.tic()
print 'Demo for {}'.format(im_path)
im = cv2.imread(im_path)
scores, boxes = im_detect(net, im)
# ignore bg
h, w, _ = im.shape
is_target = len(t_cls) > 0
for cls_ind, cls in enumerate(classes[1:]):
if is_target and cls not in t_cls:
continue
cls_ind += 1
cls_scores = scores[:, cls_ind]
order = cls_scores.argsort()[::-1]
obj_ind = order[0]
score = cls_scores[obj_ind]
bbox = boxes[obj_ind, 4 * cls_ind:4 * (cls_ind + 1)]
bbox = [int(b) for b in bbox]
bbox = [x1, y1, x2, y2]
cls = classes[cls_ind]
p1 = (bbox[0], bbox[1])
p2 = (bbox[2], bbox[3])
cv2.rectangle(im, p1, p2, (38, 231, 16), 2)
p3 = (bbox[0], bbox[1] - 5)
cv2.putText(im, '{:s} {:s}'.format(cls, str(score),), p3, \
cv2.FONT_HERSHEY_SIMPLEX, .56, (123, 19, 208), 1)
im_name = im_path.rsplit("/", 1)[1]
im_name = im_name.rsplit(".", 1)[0]
if out_dire and len(out_dire) > 0:
im_path2 = out_dire + im_name + im_ext
cv2.imwrite(im_path2, im)
else:
cv2.imshow(im_name, im)
cv2.waitKey(1)
cv2.destroyAllWindows()
total_time = timer.toc(average=False)
print "Detection took %ss for %s object proposals" \
% (total_time, boxes.shape[0])
def _demo4video(net, im, classes, t_cls, NMS_THRESH=0.3, CONF_THRESH=0.8):
"""Detect object classes in an image using pre-computed object proposals."""
timer = Timer()
timer.tic()
scores, boxes = im_detect(net, im)
# ignore bg
is_target = len(t_cls) > 0
for cls_ind, cls in enumerate(classes[1:]):
if is_target and cls not in t_cls:
continue
cls_ind += 1
cls_boxes = boxes[:, 4 * cls_ind:4 * (cls_ind + 1)]
cls_scores = scores[:, cls_ind]
dets = np.hstack(
(cls_boxes, cls_scores[:, np.newaxis])).astype(np.float32)
keep = nms(dets, NMS_THRESH)
dets = dets[keep, :]
# draw bboxes
inds = np.where(dets[:, -1] >= CONF_THRESH)[0]
if len(inds) == 0:
cv2.imshow("frame", im)
return
for i in inds:
bbox = dets[i, :4]
bbox = [int(b) for b in bbox]
score = dets[i, -1]
# bbox
x1 = bbox[0]
y1 = bbox[1]
x2 = bbox[2]
y2 = bbox[3]
# draw
p1 = (x1, y1)
p2 = (x2, y2)
cv2.rectangle(im, p1, p2, (38, 231, 16), 2)
p3 = (x1, y1 - 3)
cv2.putText(im, '{:s} {:.3f}'.format(cls, score), p3, \
cv2.FONT_HERSHEY_SIMPLEX, .36, (23, 119, 188))
# show for all categories
cv2.imshow("frame", im)
total_time = timer.toc(average=False)
print "Detection took %ss for %s object proposals" % (total_time,
boxes.shape[0])
def _image2bbox(net, im, classes, t_cls, NMS_THRESH=0.3, CONF_THRESH=0.8):
"""
Detect object classes in an image using pre-computed object proposals.
>= threshold (maybe empty)
"""
timer = Timer()
timer.tic()
scores, boxes, = im_detect(net, im)
# ignore bg
bboxes = []
is_target = len(t_cls) > 0
for cls_ind, cls in enumerate(classes[1:]):
if is_target and cls not in t_cls:
continue
cls_ind += 1
cls_boxes = boxes[:, 4 * cls_ind:4 * (cls_ind + 1)]
cls_scores = scores[:, cls_ind]
dets = np.hstack(
(cls_boxes, cls_scores[:, np.newaxis])).astype(np.float32)
keep = nms(dets, NMS_THRESH)
dets = dets[keep, :]
# draw bboxes
inds = np.where(dets[:, -1] >= CONF_THRESH)[0]
if len(inds) == 0:
break
for i in inds:
bbox = dets[i, :4]
bbox = [int(b) for b in bbox]
score = dets[i, -1]
bboxes.append((bbox, score, cls))
total_time = timer.toc(average=False)
print "Detection took %ss for %s object proposals" % (total_time,
boxes.shape[0])
return bboxes
def _demo4image2file_top1(net, im_path, fhd, classes, t_cls):
"""
Detect object classes in an image using pre-computed object proposals.
And write the results of bboxes into file by file handler `fhd`
Format:
im_path [[score bbox cls] ...]
"""
timer = Timer()
timer.tic()
print 'Demo for {}'.format(im_path)
im = cv2.imread(im_path)
scores, boxes = im_detect(net, im)
# ignore bg
h, w, _ = im.shape
info = im_path.strip()
is_target = len(t_cls) > 0
for cls_ind, cls in enumerate(classes[1:]):
if is_target and cls not in t_cls:
continue
cls_ind += 1
cls_scores = scores[:, cls_ind]
order = cls_scores.argsort()[::-1]
obj_ind = order[0]
score = cls_scores[obj_ind]
score = str(score)
bbox = boxes[obj_ind, 4 * cls_ind:4 * (cls_ind + 1)]
bbox = [int(b) for b in bbox]
bbox = [str(b) for b in bbox]
bbox = " ".join(bbox).strip()
info = info + " " + score + " " + bbox + " " + cls
# for all categories in one line
fhd.write(info.strip() + "\n")
total_time = timer.toc(average=False)
print "Detection took %ss for %s object proposals" \
% (total_time, boxes.shape[0])
def _image2bbox_top1_cond(net,
im,
classes,
t_cls,
cond_bbox,
NMS_THRESH=0.3,
CONF_THRESH=0.8):
"""
Detect object classes in an image using pre-computed object proposals.
"""
timer = Timer()
timer.tic()
scores, boxes = im_detect(net, im)
# ignore bg
bboxes = []
is_target = len(t_cls) > 0
for cls_ind, cls in enumerate(classes[1:]):
if is_target and cls not in t_cls:
continue
cls_ind += 1
cls_scores = scores[:, cls_ind]
order = cls_scores.argsort()[::-1]
for obj_ind in order:
score = cls_scores[obj_ind]
bbox = boxes[obj_ind, 4 * cls_ind:4 * (cls_ind + 1)]
bbox = [int(b) for b in bbox]
if _within_bbox(cond_bbox, bbox):
bboxes.append((bbox, score, cls))
break
total_time = timer.toc(average=False)
print "Detection took %ss for %s object proposals" % (total_time,
boxes.shape[0])
return bboxes
print "t_cls:", t_cls
print "classes:", classes
print "im_path:", im_path
print "out_dire:", out_dire
time.sleep(2)
return net, args, classes, im_path, out_dire, out_file, t_cls, im_ext
if __name__ == '__main__':
"""Demo for human or torso detection"""
net, args, classes, im_path, out_dire, out_file, t_cls, im_ext = _init_parse()
# Warmup on a dummy image
im = 128 * np.ones((300, 500, 3), dtype=np.uint8)
for i in xrange(2):
im_detect(net, im)
# #####################################################
# begin
# #####################################################
if args.is_video != 0:
print "\n\nDection of Person | Torso in Video"
time.sleep(2)
pose4video(net, classes, t_cls)
else:
print "\n\nDection of Person | Torso in images"
time.sleep(2)
pose4images_top1(net, classes, im_path, t_cls, out_dire, out_file)
print "\n\nDetection has been done.\n\n"
