def detect(interpreter,
input_blob,
output_blob,
image,
threshold=0.4,
nms_iou=0.5):
interpreter.set_tensor(input_blob[0]['index'], image)
interpreter.invoke()
# hm, box, landmark = outputs['1028'], outputs['1029'], outputs['1027']
lm = interpreter.get_tensor(output_blob[0]['index']).transpose(
(0, 3, 1, 2)) # 1,h,w,10
box = interpreter.get_tensor(output_blob[1]['index']).transpose(
(0, 3, 1, 2)) # 1,h,w,4
hm = interpreter.get_tensor(output_blob[2]['index']).transpose(
(0, 3, 1, 2)) # 1,1,h,w
x = torch.from_numpy(hm).clone()
y = torch.from_numpy(box).clone()
z = torch.from_numpy(lm).clone()
for var in [x, y, z]:
if var.shape[1] == 1:
hm = var
elif var.shape[1] == 4:
box = var
elif var.shape[1] == 10:
landmark = var
hm_pool = F.max_pool2d(hm, 3, 1, 1)
scores, indices = ((hm == hm_pool).float() * hm).view(1,
-1).cpu().topk(1000)
hm_height, hm_width = hm.shape[2:]
scores = scores.squeeze()
indices = indices.squeeze()
ys = list((indices / hm_width).int().data.numpy())
xs = list((indices % hm_width).int().data.numpy())
scores = list(scores.data.numpy())
box = box.cpu().squeeze().data.numpy()
landmark = landmark.cpu().squeeze().data.numpy()
stride = 4
objs = []
for cx, cy, score in zip(xs, ys, scores):
if score < threshold:
break
x, y, r, b = box[:, cy, cx]
xyrb = (np.array([cx, cy, cx, cy]) + [-x, -y, r, b]) * stride
x5y5 = landmark[:, cy, cx]
x5y5 = (common.exp(x5y5 * 4) + ([cx] * 5 + [cy] * 5)) * stride
box_landmark = list(zip(x5y5[:5], x5y5[5:]))
objs.append(
common.BBox(0, xyrb=xyrb, score=score, landmark=box_landmark))
return nms(objs, iou=nms_iou)
def detect_images_giou_with_netout(output_hm,
output_tlrb,
output_landmark,
threshold=0.4,
ibatch=0):
stride = 4
_, num_classes, hm_height, hm_width = output_hm.shape
hm = output_hm[ibatch].reshape(1, num_classes, hm_height, hm_width)
tlrb = output_tlrb[ibatch].cpu().data.numpy().reshape(
1, num_classes * 4, hm_height, hm_width)
landmark = output_landmark[ibatch].cpu().data.numpy().reshape(
1, num_classes * 10, hm_height, hm_width)
nmskey = _nms(hm, 3)
kscore, kinds, kcls, kys, kxs = _topk(nmskey, 2000)
kys = kys.cpu().data.numpy().astype(np.int)
kxs = kxs.cpu().data.numpy().astype(np.int)
kcls = kcls.cpu().data.numpy().astype(np.int)
key = [[], [], [], []]
for ind in range(kscore.shape[1]):
score = kscore[0, ind]
if score > threshold:
key[0].append(kys[0, ind])
key[1].append(kxs[0, ind])
key[2].append(score)
key[3].append(kcls[0, ind])
imboxs = []
if key[0] is not None and len(key[0]) > 0:
ky, kx = key[0], key[1]
classes = key[3]
scores = key[2]
for i in range(len(kx)):
class_ = classes[i]
cx, cy = kx[i], ky[i]
x1, y1, x2, y2 = tlrb[0, class_ * 4:(class_ + 1) * 4, cy, cx]
x1, y1, x2, y2 = (np.array([cx, cy, cx, cy]) +
np.array([-x1, -y1, x2, y2])) * stride
x5y5 = landmark[0, 0:10, cy, cx]
x5y5 = np.array(common.exp(x5y5 * 4))
x5y5 = (x5y5 + np.array([cx] * 5 + [cy] * 5)) * stride
boxlandmark = list(zip(x5y5[:5], x5y5[5:]))
imboxs.append(
common.BBox(label=str(class_),
xyrb=common.floatv([x1, y1, x2, y2]),
score=scores[i].item(),
landmark=boxlandmark))
return imboxs
def detect(model,
image,
threshold=0.4,
nms_iou=0.5) -> typing.List[common.BBox]:
mean = [0.408, 0.447, 0.47]
std = [0.289, 0.274, 0.278]
image = common.pad(image)
image = ((image / 255.0 - mean) / std).astype(np.float32)
image = image.transpose(2, 0, 1)
torch_image = torch.from_numpy(image)[None]
if HAS_CUDA:
torch_image = torch_image.cuda()
hm, box, landmark = model(torch_image)
hm_pool = F.max_pool2d(hm, 3, 1, 1)
scores, indices = ((hm == hm_pool).float() * hm).view(1,
-1).cpu().topk(1000)
hm_height, hm_width = hm.shape[2:]
scores = scores.squeeze()
indices = indices.squeeze()
ys = list((indices / hm_width).int().data.numpy())
xs = list((indices % hm_width).int().data.numpy())
scores = list(scores.data.numpy())
box = box.cpu().squeeze().data.numpy()
landmark = landmark.cpu().squeeze().data.numpy()
stride = 4
objs = []
for cx, cy, score in zip(xs, ys, scores):
if score < threshold:
break
x, y, r, b = box[:, cy, cx]
xyrb = (np.array([cx, cy, cx, cy]) + [-x, -y, r, b]) * stride
x5y5 = landmark[:, cy, cx]
x5y5 = (common.exp(x5y5 * 4) + ([cx] * 5 + [cy] * 5)) * stride
box_landmark = list(zip(x5y5[:5], x5y5[5:]))
objs.append(
common.BBox(0, xyrb=xyrb, score=score, landmark=box_landmark))
return nms(objs, iou=nms_iou)
def detect(exec_net, input_blob, image, threshold=0.4, nms_iou=0.5):
outputs = exec_net.infer(inputs={input_blob: image})
# print('outputs:', outputs)
# print('outputs[\'Sigmoid_526\'].shape:', outputs['Sigmoid_526'].shape)
# print('outputs[\'Exp_527\'].shape:', outputs['Exp_527'].shape)
# print('outputs[\'Conv_525\'].shape:', outputs['Conv_525'].shape)
hm, box, landmark = outputs['Sigmoid_526'], outputs['Exp_527'], outputs[
'Conv_525']
hm = torch.from_numpy(hm).clone()
box = torch.from_numpy(box).clone()
landmark = torch.from_numpy(landmark).clone()
hm_pool = F.max_pool2d(hm, 3, 1, 1)
scores, indices = ((hm == hm_pool).float() * hm).view(1,
-1).cpu().topk(1000)
hm_height, hm_width = hm.shape[2:]
scores = scores.squeeze()
indices = indices.squeeze()
ys = list((indices / hm_width).int().data.numpy())
xs = list((indices % hm_width).int().data.numpy())
scores = list(scores.data.numpy())
box = box.cpu().squeeze().data.numpy()
landmark = landmark.cpu().squeeze().data.numpy()
stride = 4
objs = []
for cx, cy, score in zip(xs, ys, scores):
if score < threshold:
break
x, y, r, b = box[:, cy, cx]
xyrb = (np.array([cx, cy, cx, cy]) + [-x, -y, r, b]) * stride
x5y5 = landmark[:, cy, cx]
x5y5 = (common.exp(x5y5 * 4) + ([cx] * 5 + [cy] * 5)) * stride
box_landmark = list(zip(x5y5[:5], x5y5[5:]))
objs.append(
common.BBox(0, xyrb=xyrb, score=score, landmark=box_landmark))
return nms(objs, iou=nms_iou)
def detect_images_giou_with_retinaface_style_eval(output_hm,
output_tlrb,
output_landmark,
threshold=0.4,
ibatch=0):
stride = 4
_, _, hm_height, hm_width = output_hm.shape
hm = output_hm[ibatch].reshape(1, 1, hm_height, hm_width)
tlrb = output_tlrb[ibatch]
landmark = output_landmark[ibatch]
area = hm_height * hm_width
keep = (hm > threshold).view(area)
indices = torch.arange(0, area)[keep]
hm = hm.view(1, area).cpu().data.numpy()
tlrb = tlrb.view(4, area).cpu().data.numpy()
landmark = landmark.view(10, area).cpu().data.numpy()
cx, cy = indices % hm_width, indices // hm_width
scores = hm[0, indices]
x1, y1, x2, y2 = tlrb[0:4, indices]
cts = np.vstack([cx, cy, cx, cy])
locs = np.vstack([-x1, -y1, x2, y2])
x1, y1, x2, y2 = (cts + locs) * stride
x5y5 = landmark[0:10, indices]
x5y5 = common.exp(x5y5 * 4)
x5y5 = (x5y5 + np.vstack([cx] * 5 + [cy] * 5)) * stride
imboxs = []
for i in range(len(indices)):
boxlandmark = list(zip(x5y5[0:5, i], x5y5[5:, i]))
imboxs.append(
common.BBox(label="facial",
xyrb=common.floatv([x1[i], y1[i], x2[i], y2[i]]),
score=scores[i],
landmark=boxlandmark))
return imboxs