def __call__(self, ori_img):
# img to tensor
assert isinstance(ori_img, np.ndarray), "input must be a numpy array!"
# 转换类型
img = ori_img.astype(np.float) / 255.
img = cv2.resize(img, self.size)
# https://blog.csdn.net/zzw000000/article/details/80320040
# 将numpy的格式转化为tensor格式
# permute(dims) 将tensor的维度换位
# unsqueeze()这个函数主要是对数据维度进行扩充。 ???
img = torch.from_numpy(img).float().permute(2, 0, 1).unsqueeze(0)
# forward
with torch.no_grad():
img = img.to(self.device)
out_boxes = self.net(img)
boxes = get_all_boxes(out_boxes,
self.conf_thresh,
self.net.num_classes,
use_cuda=self.use_cuda)[0]
boxes = nms(boxes, self.nms_thresh)
# print(boxes)
# plot boxes
if self.is_plot:
return self.plot_bbox(ori_img, boxes)
if len(boxes) == 0:
return None, None, None
height, width = ori_img.shape[:2]
boxes = np.vstack(boxes)
bbox = np.empty_like(boxes[:, :4])
if self.is_xywh:
# bbox x y w h
bbox[:, 0] = boxes[:, 0] * width
bbox[:, 1] = boxes[:, 1] * height
bbox[:, 2] = boxes[:, 2] * width
bbox[:, 3] = boxes[:, 3] * height
else:
# bbox xmin ymin xmax ymax
bbox[:, 0] = (boxes[:, 0] - boxes[:, 2] / 2.0) * width
bbox[:, 1] = (boxes[:, 1] - boxes[:, 3] / 2.0) * height
bbox[:, 2] = (boxes[:, 0] + boxes[:, 2] / 2.0) * width
bbox[:, 3] = (boxes[:, 1] + boxes[:, 3] / 2.0) * height
cls_conf = boxes[:, 5]
cls_ids = boxes[:, 6]
# print(cls_ids)
return bbox, cls_conf, cls_ids
def __call__(self, ori_img):
# img to tensor
assert isinstance(ori_img, np.ndarray), "input must be a numpy array!"
img = ori_img.astype(np.float) / 255.
img = cv2.resize(img, self.size)
img = torch.from_numpy(img).float().permute(2, 0, 1).unsqueeze(0)
# forward
with torch.no_grad():
img = img.to(self.device)
out_boxes = self.net(img)
boxes = get_all_boxes(out_boxes, self.conf_thresh,
self.net.num_classes, self.use_cuda)[0]
boxes = nms(boxes, self.nms_thresh)
print(type(boxes))
re_boxes = boxes
# print(boxes)
# plot boxes
if self.is_plot:
return self.plot_bbox(ori_img, boxes)
if len(boxes) == 0:
return None, None, None
height, width = ori_img.shape[:2]
boxes = np.vstack(boxes)
bbox = np.empty_like(boxes[:, :4])
if self.is_xywh:
# bbox x y w h
bbox[:, 0] = boxes[:, 0] * width
bbox[:, 1] = boxes[:, 1] * height
bbox[:, 2] = boxes[:, 2] * width
bbox[:, 3] = boxes[:, 3] * height
else:
# bbox xmin ymin xmax ymax
bbox[:, 0] = (boxes[:, 0] - boxes[:, 2] / 2.0) * width
bbox[:, 1] = (boxes[:, 1] - boxes[:, 3] / 2.0) * height
bbox[:, 2] = (boxes[:, 0] + boxes[:, 2] / 2.0) * width
bbox[:, 3] = (boxes[:, 1] + boxes[:, 3] / 2.0) * height
cls_conf = boxes[:, 5]
cls_ids = boxes[:, 6]
# print(bbox[:,0], bbox[:,1], bbox[:,2], bbox[:,3])
# return bbox, cls_conf, cls_ids
return re_boxes, cls_conf, cls_ids
def yolo_crop(from_mat):
height, width = from_mat.shape[:2]
original_image = from_mat
# original_image = cv2.cvtColor(original_image, cv2.COLOR_BGR2RGB)
original_image_size = original_image.shape[:2]
image_data = utils.image_preporcess(np.copy(original_image), [input_size, input_size])
image_data = image_data[np.newaxis, ...]
pred_sbbox, pred_mbbox, pred_lbbox = sess.run(
[return_tensors[1], return_tensors[2], return_tensors[3]],
feed_dict={return_tensors[0]: image_data})
# print(pred_sbbox.shape, pred_mbbox.shape, pred_lbbox.shape)
pred_bbox = np.concatenate([np.reshape(pred_sbbox, (-1, 5 + num_classes)),
np.reshape(pred_mbbox, (-1, 5 + num_classes)),
np.reshape(pred_lbbox, (-1, 5 + num_classes))], axis=0)
# bboxes: (xmin, ymin, xmax, ymax, score, class)
bboxes = utils.postprocess_boxes(pred_bbox, original_image_size, input_size, 0.3)
bboxes = utils.nms(bboxes, 0.45, method='nms')
# image = utils.draw_bbox(original_image, bboxes)
# img_arr = image
# image = Image.fromarray(image)
for (x, y, xm, ym, score, cls) in bboxes:
low_res = False
out_bounds = False
relaxed_fit = False
# BB new crop 4
if False:
w = xm - x
h = ym - y
x_mid = (x + xm) * 0.5
y_mid = (ym - (xm - x) * 0.5)
s = w
crop_x = int(x_mid - s * 1.2)
crop_x2 = crop_x + int(s * 2.4)
crop_y2 = int(y_mid + s * 1.1)
crop_y = crop_y2 - int(s * 2.4)
if crop_x < 0:
crop_x = 0
out_bounds = True
if crop_y < 0:
crop_y = 0
out_bounds = True
if crop_x2 > width:
crop_x2 = width
out_bounds = True
if crop_y2 > height:
crop_y2 = height
out_bounds = True
# cropped = from_mat[crop_y:crop_y2, crop_x:crop_x2]
elif True: # halfway between bbn4 and no expanding
w = xm - x
h = ym - y
x_mid = (x + xm) * 0.5
y_mid = (ym - (xm - x) * 0.5)
s = w
crop_x = int(x_mid - s * 0.9)
crop_x2 = crop_x + int(s * 1.8)
crop_y2 = int(y_mid + s * 0.825)
crop_y = crop_y2 - int(s * 1.8)
if crop_x < 0:
crop_x = 0
out_bounds = True
if crop_y < 0:
crop_y = 0
out_bounds = True
if crop_x2 > width:
crop_x2 = width
out_bounds = True
if crop_y2 > height:
crop_y2 = height
out_bounds = True
else:
print("ERRROR ERROR NO CROP SELECTED")
exit()
# need to return to match fa.face_detector.detect_from_image
# [[min_x, min_y, max_x, max_y, prob]...]
# basic crop
# yield [x, y, xm, ym, score]
# crop above
yield [crop_x, crop_y, crop_x2, crop_y2, score]