def __call__(self, src, bbox):
# resize with random interpolation
h, w, _ = src.shape
interp = np.random.randint(1, 5)
scale = 1.2
src = timage.imresize(src,
int(self._width * scale),
int(self._height * scale),
interp=interp)
bbox = tbbox.resize(
bbox, (w, h),
(int(self._width * scale), int(self._height * scale)))
# random color jittering
img = experimental.image.random_color_distort(src)
# random cropping
h, w, _ = img.shape
bbox, crop = random_crop_with_constraints(bbox, (w, h),
self._height,
self._width,
min_scale=0.95,
max_scale=1.05,
max_trial=50)
x0, y0, w, h = crop
img = mx.image.fixed_crop(img, x0, y0, w, h)
h, w, _ = img.shape
img, flips = timage.random_flip(img, px=0.5)
bbox = tbbox.flip(bbox, (w, h), flip_x=flips[0])
# resize with random interpolation
h, w, _ = img.shape
interp = np.random.randint(1, 5)
img = timage.imresize(img, self._width, self._height, interp=interp)
bbox = tbbox.resize(bbox, (w, h), (self._width, self._height))
# random horizontal flip
h, w, _ = img.shape
img, flips = timage.random_flip(img, px=0.5)
bbox = tbbox.flip(bbox, (w, h), flip_x=flips[0])
# to tensor
img = mx.nd.image.to_tensor(img)
img = mx.nd.image.normalize(img, mean=self._mean, std=self._std)
if self._anchors is None:
return img, bbox.astype(img.dtype)
# generate training target so cpu workers can help reduce the workload on gpu
gt_bboxes = mx.nd.array(bbox[np.newaxis, :, :4])
gt_ids = mx.nd.array(bbox[np.newaxis, :, 4:5])
cls_targets, box_targets, _ = self._target_generator(
self._anchors, None, gt_bboxes, gt_ids)
return img, cls_targets[0], box_targets[0]
def __call__(self, src, bbox):
img = src
# resize with random interpolation
h, w, _ = img.shape
interp = np.random.randint(1, 5)
img = timage.imresize(img, self._width, self._height, interp=interp)
bbox = tbbox.resize(bbox, (w, h), (self._width, self._height))
if not self._val:
h, w, _ = img.shape
img, flips = timage.random_flip(img, px=0.5)
bbox = tbbox.flip(bbox, (w, h), flip_x=flips[0])
# to tensor
img = mx.nd.image.to_tensor(img)
img[0, :, :] = mx.nd.subtract(img[0, :, :] * 256, self._mean[0])
img[1, :, :] = mx.nd.subtract(img[1, :, :] * 256, self._mean[1])
img[2, :, :] = mx.nd.subtract(img[2, :, :] * 256, self._mean[2])
if self._anchors is None:
return img, bbox.astype(img.dtype)
# generate training target so cpu workers can help reduce the workload on gpu
gt_bboxes = mx.nd.array(bbox[np.newaxis, :, :4])
gt_ids = mx.nd.array(bbox[np.newaxis, :, 4:5])
cls_targets, box_targets, _ = self._target_generator(
self._anchors, None, gt_bboxes, gt_ids)
return img, cls_targets[0], box_targets[0]
def __call__(self, src, label):
"""Apply transform to training image/label."""
# random color jittering
img = random_color_distort(src)
# random cropping #! keep aspect ration = 1
h, w, _ = img.shape
bbox, crop = random_crop_with_constraints(label, (w, h))
x0, y0, w, h = crop
img = mx.image.fixed_crop(img, x0, y0, w, h)
# resize with random interpolation
h, w, _ = img.shape
interp = np.random.randint(0, 5)
img = gimage.imresize(img, self._width, self._height, interp=interp)
bbox = gbbox.resize(bbox, (w, h), (self._width, self._height))
# random horizontal flip
h, w, _ = img.shape
img, flips = gimage.random_flip(img, px=0.5)
bbox = gbbox.flip(bbox, (w, h), flip_x=flips[0])
# to tensor
img = mx.nd.image.to_tensor(img)
img = mx.nd.image.normalize(img, mean=self._mean, std=self._std)
if self._anchors is None:
return img, bbox
# generate training target so cpu workers can help reduce the workload on gpu
gt_bboxes = mx.nd.array(bbox[:, :4]).expand_dims(0)
gt_ids = mx.nd.zeros((1, gt_bboxes.shape[1], 1), dtype=gt_bboxes.dtype)
cls_targets, box_targets, _ = self._target_generator(
self._anchors, None, gt_bboxes, gt_ids)
return img, cls_targets[0], box_targets[0]
def __call__(self, src, label):
"""Apply transform to training image/label."""
# random color jittering
img = experimental.image.random_color_distort(src)
# random expansion with prob 0.5
if np.random.uniform(0, 1) > 0.5:
img, expand = timage.random_expand(
img, fill=[m * 255 for m in self._mean])
bbox = tbbox.translate(label,
x_offset=expand[0],
y_offset=expand[1])
else:
img, bbox = img, label
# random cropping
h, w, _ = img.shape
bbox, crop = experimental.bbox.random_crop_with_constraints(
bbox, (w, h))
x0, y0, w, h = crop
img = mx.image.fixed_crop(img, x0, y0, w, h)
# resize with random interpolation
h, w, _ = img.shape
interp = np.random.randint(0, 5)
img = timage.imresize(img, self._width, self._height, interp=interp)
bbox = tbbox.resize(bbox, (w, h), (self._width, self._height))
# random horizontal flip
h, w, _ = img.shape
img, flips = timage.random_flip(img, px=0.5)
bbox = tbbox.flip(bbox, (w, h), flip_x=flips[0])
if (self._bilateral_kernel_size is not None) and (
self._sigma_vals is not None) or self._grayscale:
img = img.asnumpy()
if (self._bilateral_kernel_size is not None) and (self._sigma_vals
is not None):
img = cv2.bilateralFilter(img, self._bilateral_kernel_size,
self._sigma_vals, self._sigma_vals)
if self._grayscale:
img = np.dot(img[..., :3], [0.299, 0.587, 0.114])
img = np.repeat(img[:, :, None], 3, axis=2)
img = nd.array(img)
# to tensor
img = mx.nd.image.to_tensor(img)
img = mx.nd.image.normalize(img, mean=self._mean, std=self._std)
if self._anchors is None:
return img, bbox.astype(img.dtype)
# generate training target so cpu workers can help reduce the workload on gpu
gt_bboxes = mx.nd.array(bbox[np.newaxis, :, :4])
gt_ids = mx.nd.array(bbox[np.newaxis, :, 4:5])
cls_targets, box_targets, _ = self._target_generator(
self._anchors, None, gt_bboxes, gt_ids)
return img, cls_targets[0], box_targets[0]
def __call__(self, src, label, segm):
"""Apply transform to training image/label."""
# resize shorter side but keep in max_size
h, w, _ = src.shape
if self._random_resize:
short = randint(self._short[0], self._short[1])
else:
short = self._short
img = timage.resize_short_within(src, short, self._max_size, interp=1)
bbox = tbbox.resize(label, (w, h), (img.shape[1], img.shape[0]))
# segm = [tmask.resize(polys, (w, h), (img.shape[1], img.shape[0])) for polys in segm]
# random horizontal flip
h, w, _ = img.shape
img, flips = timage.random_flip(img, px=0.5)
bbox = tbbox.flip(bbox, (w, h), flip_x=flips[0])
# segm = [tmask.flip(polys, (w, h), flip_x=flips[0]) for polys in segm]
# gt_masks (n, im_height, im_width) of uint8 -> float32 (cannot take uint8)
# masks = [mx.nd.array(tmask.to_mask(polys, (w, h))) for polys in segm]
masks = cocomask.decode(segm) # hxwxn
mask_list = []
for i in range(masks.shape[-1]):
mask = cv2.resize(masks[:,:,i], (img.shape[1],img.shape[0]),
interpolation=cv2.INTER_NEAREST)
mask_list.append(mx.nd.array(mask))
# n * (im_height, im_width) -> (n, im_height, im_width)
masks = mx.nd.stack(*mask_list, axis=0)
if flips[0]:
masks = mx.nd.flip(masks, axis=2)
# to tensor
img = mx.nd.image.to_tensor(img)
img = mx.nd.image.normalize(img, mean=self._mean, std=self._std)
if self._anchors is None:
return img, bbox.astype(img.dtype), masks
# generate RPN target so cpu workers can help reduce the workload
# feat_h, feat_w = (img.shape[1] // self._stride, img.shape[2] // self._stride)
gt_bboxes = mx.nd.array(bbox[:, :4])
if self._multi_stage:
oshapes = []
anchor_targets = []
for feat_sym in self._feat_sym:
oshapes.append(feat_sym.infer_shape(data=(1, 3, img.shape[1], img.shape[2]))[1][0])
for anchor, oshape in zip(self._anchors, oshapes):
anchor = anchor[:, :, :oshape[2], :oshape[3], :].reshape((-1, 4))
anchor_targets.append(anchor)
anchor_targets = mx.nd.concat(*anchor_targets, dim=0)
cls_target, box_target, box_mask = self._target_generator(
gt_bboxes, anchor_targets, img.shape[2], img.shape[1])
else:
oshape = self._feat_sym.infer_shape(data=(1, 3, img.shape[1], img.shape[2]))[1][0]
anchor = self._anchors[:, :, :oshape[2], :oshape[3], :].reshape((-1, 4))
cls_target, box_target, box_mask = self._target_generator(
gt_bboxes, anchor, img.shape[2], img.shape[1])
return img, bbox.astype(img.dtype), masks, cls_target, box_target, box_mask
def __call__(self, src, label):
"""Apply transform to validation image/label."""
# resize
h, w, _ = src.shape
img = timage.imresize(src, self._width, self._height, interp=9)
bbox = tbbox.resize(label, in_size=(w, h), out_size=(self._width, self._height))
img = mx.nd.image.to_tensor(img)
img = mx.nd.image.normalize(img, mean=self._mean, std=self._std)
return img, bbox.astype(img.dtype)
def __call__(self, src, label):
"""Apply transform to training image/label."""
"""color distort"""
# img = random_color_distort(src)
# print("previous label shape = ", label.shape)
target = np.zeros(shape=(label.shape[0], ))
"""Pyramid Anchor sampling"""
img, boxes, label = self.random_baiducrop(src, label[:, :4], target)
# print("label shape = ", label.shape)
# print('boxes shape =', boxes.shape)
bbox = boxes
# img = mx.nd.array(img)
"""color distort"""
img = mx.nd.array(img)
img = random_color_distort(img)
# """random crop, keep aspect ration=1"""
# h, w, _ = img.shape
# bbox, crop_size = random_crop_with_constraints(label, (w, h))
# x_offset, y_offset, new_width, new_height = crop_size
# img = mx.image.fixed_crop(img, x_offset, y_offset, new_width, new_height)
"""resize with random interpolation"""
h, w, _ = img.shape
interp = np.random.randint(0, 5)
img = gimage.imresize(img, self._width, self._height, interp=interp)
bbox = gbbox.resize(bbox, (w, h), (self._width, self._height))
"""random horizontal flip"""
h, w, _ = img.shape
img, flips = gimage.random_flip(img, px=0.5)
bbox = gbbox.flip(bbox, (w, h), flip_x=flips[0])
"""To Tensor & Normalization"""
img = mx.nd.image.to_tensor(img)
img = mx.nd.image.normalize(img, mean=self._mean, std=self._std)
if self._anchors is None:
return img, bbox
# @TODO: generating training target so cpu workers can help reduce the workload on gpu
face_anchors, head_anchors, body_anchors = self._anchors
gt_bboxes = mx.nd.array(bbox[:, :4]).expand_dims(0)
gt_ids = mx.nd.zeros((1, gt_bboxes.shape[1], 1), dtype=gt_bboxes.dtype)
face_cls_targets, face_box_targets, _ = self._target_generator(
face_anchors, None, gt_bboxes, gt_ids)
head_cls_targets, head_box_targets, _ = self._target_generator(
head_anchors, None, gt_bboxes, gt_ids)
body_cls_targets, body_box_targets, _ = self._target_generator(
body_anchors, None, gt_bboxes, gt_ids)
return img, \
face_cls_targets[0], head_cls_targets[0], body_cls_targets[0], \
face_box_targets[0], head_box_targets[0], body_box_targets[0]
def transform_gt_bbox(img_path, model, bbox):
if model == 'yolo':
short, max_size = 416, 1024
if model == 'rcnn':
short, max_size = 600, 1000
img = mx.image.imread(img_path)
h, w, _ = img.shape
resized_img = timage.resize_short_within(img, short, max_size)
bbox = tbbox.resize(bbox, (w, h),
(resized_img.shape[1], resized_img.shape[0]))
return bbox
def process_frame(image, net, ctx):
# currently only supports batch size 1 todo
image = np.squeeze(image)
image = mx.nd.array(image, dtype='uint8')
x, _ = transform_test(image, 600, max_size=1000)
x = x.copyto(ctx[0])
# get prediction results
ids, scores, bboxes = net(x)
oh, ow, _ = image.shape
_, _, ih, iw = x.shape
bboxes[0] = tbbox.resize(bboxes[0], in_size=(iw, ih), out_size=(ow, oh))
return bboxes[0].asnumpy(), scores[0].asnumpy(), ids[0].asnumpy()
def __call__(self, src, label):
"""Apply transform to training image/label."""
# random color jittering
img = experimental.image.random_color_distort(src)
# random expansion with prob 0.5
if np.random.uniform(0, 1) > 0.5:
img, expand = timage.random_expand(
img, fill=[m * 255 for m in self._mean])
bbox = tbbox.translate(label,
x_offset=expand[0],
y_offset=expand[1])
else:
img, bbox = img, label
# random cropping
h, w, _ = img.shape
bbox, crop = experimental.bbox.random_crop_with_constraints(
bbox, (w, h))
x0, y0, w, h = crop
img = mx.image.fixed_crop(img, x0, y0, w, h)
# resize with random interpolation
h, w, _ = img.shape
interp = np.random.randint(0, 5)
img = timage.imresize(img, self._width, self._height, interp=interp)
bbox = tbbox.resize(bbox, (w, h), (self._width, self._height))
# random horizontal flip
h, w, _ = img.shape
img, flips = timage.random_flip(img, px=0.5)
bbox = tbbox.flip(bbox, (w, h), flip_x=flips[0])
# to tensor
img = mx.nd.image.to_tensor(img)
img = mx.nd.image.normalize(img, mean=self._mean, std=self._std)
if self._anchors is None:
return img, bbox.astype(img.dtype)
# generate training target so cpu workers can help reduce the workload on gpu
gt_bboxes = mx.nd.array(bbox[np.newaxis, :, :4])
gt_ids = mx.nd.array(bbox[np.newaxis, :, 4:5])
anchor_cls_targets, anchor_box_targets, _ = self._target_generator(
self._anchors, None, gt_bboxes, gt_ids)
anchor_cls_targets = mx.nd.where(anchor_cls_targets > 0,
mx.nd.ones_like(anchor_cls_targets),
anchor_cls_targets)
# positive anchor is 1, negative anchor is 0 and ignored is -1.
return img, anchor_cls_targets[0], anchor_box_targets[
0], bbox[:, :5].astype(img.dtype)
def __call__(self, src, bbox):
"""Apply transform to training image/label."""
if not self._val:
# random color jittering
src = experimental.image.random_color_distort(src)
img = src
# random cropping
h, w, _ = img.shape
bbox, crop = random_crop_with_constraints(bbox, (w, h),
self._height,
self._width,
min_scale=0.9,
max_scale=1,
max_trial=50)
x0, y0, w, h = crop
img = mx.image.fixed_crop(img, x0, y0, w, h)
# resize with random interpolation
h, w, _ = img.shape
interp = np.random.randint(1, 5)
img = timage.imresize(img, self._width, self._height, interp=interp)
bbox = tbbox.resize(bbox, (w, h), (self._width, self._height))
if not self._val:
# random horizontal flip
h, w, _ = img.shape
img, flips = timage.random_flip(img, px=0.5)
bbox = tbbox.flip(bbox, (w, h), flip_x=flips[0])
# to tensor
img = mx.nd.image.to_tensor(img)
img[0, :, :] = mx.nd.subtract(img[0, :, :] * 256, self._mean[0])
img[1, :, :] = mx.nd.subtract(img[1, :, :] * 256, self._mean[1])
img[2, :, :] = mx.nd.subtract(img[2, :, :] * 256, self._mean[2])
if self._anchors is None:
return img, bbox.astype(img.dtype)
# generate training target so cpu workers can help reduce the workload on gpu
gt_bboxes = mx.nd.array(bbox[np.newaxis, :, :4])
gt_ids = mx.nd.array(bbox[np.newaxis, :, 4:5])
cls_targets, box_targets, _ = self._target_generator(
self._anchors, None, gt_bboxes, gt_ids)
return img, cls_targets[0], box_targets[0]
def __call__(self, src, label):
"""Apply transform to training image/label."""
# random color jittering
img = experimental.image.random_color_distort(src)
# random expansion with prob 0.5
if np.random.uniform(0, 1) > 0.5:
img, expand = timage.random_expand(img, fill=[m * 255 for m in self._mean])
bbox = tbbox.translate(label, x_offset=expand[0], y_offset=expand[1])
else:
img, bbox = img, label
# random cropping
h, w, _ = img.shape
bbox, crop = experimental.bbox.random_crop_with_constraints(bbox, (w, h))
x0, y0, w, h = crop
img = mx.image.fixed_crop(img, x0, y0, w, h)
# resize with random interpolation
h, w, _ = img.shape
interp = np.random.randint(0, 5)
img = timage.imresize(img, self._width, self._height, interp=interp)
bbox = tbbox.resize(bbox, (w, h), (self._width, self._height))
# random horizontal flip
h, w, _ = img.shape
img, flips = timage.random_flip(img, px=0.5)
bbox = tbbox.flip(bbox, (w, h), flip_x=flips[0])
# to tensor
img = mx.nd.image.to_tensor(img)
img = mx.nd.image.normalize(img, mean=self._mean, std=self._std)
if self._anchors is None:
return img, bbox.astype(img.dtype)
#如果有anchors的输入，则执行下面的运算。计算以前的格式是：
#gt_bboxes里面是一个图像上的box位置，实际像素点位 当前图片里面的框框个数x4
#gt_ids里面是对应图像上的label，当前图片里面的框框个数x1
#下面为batch_size腾出空间
gt_bboxes = mx.nd.array(bbox[np.newaxis, :, :4])
gt_ids = mx.nd.array(bbox[np.newaxis, :, 4:5])
cls_targets, box_targets, _ = self._target_generator(
self._anchors, None, gt_bboxes, gt_ids)
return img, cls_targets[0], box_targets[0]
def __call__(self, src, label):
"""Apply transform to validation image/label."""
# resize with random interpolation
h, w, _ = src.shape
img = timage.imresize(src, self._width, self._height, interp=9)
bbox = tbbox.resize(label, (w, h), (self._width, self._height))
# to tensor
img = mx.nd.image.to_tensor(img)
img = mx.nd.image.normalize(img, mean=self._mean, std=self._std)
if self._anchors is None:
return img, bbox.astype(img.dtype)
# generate training target so cpu workers can help reduce the workload on gpu
gt_bboxes = mx.nd.array(bbox[np.newaxis, :, :4])
gt_ids = mx.nd.array(bbox[np.newaxis, :, 4:5])
cls_targets, box_targets, _ = self._target_generator(
self._anchors, None, gt_bboxes, gt_ids)
return img, cls_targets[0], box_targets[0]
def new_trainloader_call(self, src, label):
'''
define a new call for trainloader by changing the data augmentation
'''
# random color jittering
img = experimental.image.random_color_distort(src)
# random expansion with prob 0.5
if np.random.uniform(0, 1) > 0.5:
img, expand = timage.random_expand(img,
fill=[m * 255 for m in self._mean])
bbox = tbbox.translate(label, x_offset=expand[0], y_offset=expand[1])
else:
img, bbox = img, label
# random cropping
h, w, _ = img.shape
bbox, crop = experimental.bbox.random_crop_with_constraints(bbox, (w, h))
x0, y0, w, h = crop
img = mx.image.fixed_crop(img, x0, y0, w, h)
# resize with random interpolation
h, w, _ = img.shape
interp = np.random.randint(0, 5)
img = timage.imresize(img, self._width, self._height, interp=interp)
bbox = tbbox.resize(bbox, (w, h), (self._width, self._height))
# to tensor
img = mx.nd.image.to_tensor(img)
img = mx.nd.image.normalize(img, mean=self._mean, std=self._std)
if self._anchors is None:
return img, bbox.astype(img.dtype)
# generate training target so cpu workers can help reduce the workload on gpu
gt_bboxes = mx.nd.array(bbox[np.newaxis, :, :4])
gt_ids = mx.nd.array(bbox[np.newaxis, :, 4:5])
cls_targets, box_targets, _ = self._target_generator(
self._anchors, None, gt_bboxes, gt_ids)
return img, cls_targets[0], box_targets[0]
def __call__(self, src, label):
"""Apply transform to validation image/label."""
# resize
h, w, _ = src.shape
img = timage.imresize(src, self._width, self._height, interp=9)
bbox = tbbox.resize(label,
in_size=(w, h),
out_size=(self._width, self._height))
if (self._bilateral_kernel_size is not None) and (
self._sigma_vals is not None) or self._grayscale:
img = img.asnumpy()
if (self._bilateral_kernel_size is not None) and (self._sigma_vals
is not None):
img = cv2.bilateralFilter(img, self._bilateral_kernel_size,
self._sigma_vals, self._sigma_vals)
if self._grayscale:
img = np.dot(img[..., :3], [0.299, 0.587, 0.114])
img = np.repeat(img[:, :, None], 3, axis=2)
img = nd.array(img)
img = mx.nd.image.to_tensor(img)
img = mx.nd.image.normalize(img, mean=self._mean, std=self._std)
return img, bbox.astype(img.dtype)
def __call__(self, src, label):
"""Apply transform to training image/label."""
# random color jittering
img = experimental.image.random_color_distort(src)
# img, bbox = img,label
# random expansion with prob 0.5
if np.random.uniform(0, 1) > 0.5:
img, expand = timage.random_expand(
img, fill=[m * 255 for m in self._mean])
bbox = tbbox.translate(label,
x_offset=expand[0],
y_offset=expand[1])
else:
img, bbox = img, label
# random cropping
h, w, _ = img.shape
bbox, crop = experimental.bbox.random_crop_with_constraints(
bbox, (w, h))
x0, y0, w, h = crop
img = mx.image.fixed_crop(img, x0, y0, w, h)
# resize with random interpolation
h, w, _ = img.shape
interp = np.random.randint(0, 5)
img = timage.imresize(img, self._width, self._height, interp=interp)
bbox = tbbox.resize(bbox, (w, h), (self._width, self._height))
# random horizontal flip
# h, w, _ = img.shape
# img, flips = timage.random_flip(img, px=0.5)
# bbox = tbbox.flip(bbox, (w, h), flip_x=flips[0])
# rabdom rotation
h, w, _ = img.shape
clockwise_rotation_num = np.random.randint(0, 4)
if clockwise_rotation_num == 0:
pass
elif clockwise_rotation_num == 1:
###顺时针90度
img = nd.transpose(img, [1, 0, 2])
img = img[:, ::-1, :]
bbox = np.array([
h - bbox[:, 3], bbox[:, 0], h - bbox[:, 1], bbox[:, 2],
bbox[:, 4], bbox[:, 5]
]).T
bbox[:, 5] = (bbox[:, 5] + 1) % 4
elif clockwise_rotation_num == 2:
##顺时针180度
img = img[::-1, ::-1, :]
bbox = np.array([
w - bbox[:, 2], h - bbox[:, 3], w - bbox[:, 0], h - bbox[:, 1],
bbox[:, 4], bbox[:, 5]
]).T
bbox[:, 5] = (bbox[:, 5] + 2) % 4
else:
# 顺时针270度
img = nd.transpose(img, [1, 0, 2])
img = img[::-1, :, :]
bbox = np.array([
bbox[:, 1], w - bbox[:, 2], bbox[:, 3], w - bbox[:, 0],
bbox[:, 4], bbox[:, 5]
]).T
bbox[:, 5] = (bbox[:, 5] + 3) % 4
# to tensor
img = mx.nd.image.to_tensor(img)
img = mx.nd.image.normalize(img, mean=self._mean, std=self._std)
if self._anchors is None:
return img, bbox.astype(img.dtype)
# generate training target so cpu workers can help reduce the workload on gpu
gt_bboxes = mx.nd.array(bbox[np.newaxis, :, :4])
gt_ids = mx.nd.array(bbox[np.newaxis, :, 4:5])
gt_ori = mx.nd.array(bbox[np.newaxis, :, 5:6])
cls_targets, box_targets, _ = self._target_generator(
self._anchors, None, gt_bboxes, gt_ids)
ori_targets, box_targets, _ = self._target_generator(
self._anchors, None, gt_bboxes, gt_ori)
return img, cls_targets[0], ori_targets[0], box_targets[0]
def network_inference(self):
# a = cv2.waitKey(0) # close window when ESC is pressed
# while a is not 27:
color_img = self.color_img
depth_image = self.depth_image
depth_height_res, depth_width_res = depth_image.shape
# It is to correct the image size to fit a perfect square
# color_img = np.zeros((640, 640, 3)).astype('uint8')
# color_img[0:479] = color_img_raw[0:479]
# color_img = color_img.astype('uint8')
# Image pre-processing
frame = mx.nd.array(cv2.cvtColor(color_img, cv2.COLOR_BGR2RGB)).astype('uint8')
frame = timage.imresize(frame, self.width, self.height, 1)
frame_tensor = mx.nd.image.to_tensor(frame)
frame_tensor = mx.nd.image.normalize(frame_tensor, mean=self.mean, std=self.std)
# with TimeIt('Obj detection time'):
# Run frame through network
class_IDs, scores, bounding_boxes = self.net(frame_tensor.expand_dims(axis=0).as_in_context(self.ctx))
# Filter bounding boxes by their scores
fbounding_boxes, fscores, fclass_IDs = self.filter_predictions(bounding_boxes, scores, class_IDs)
# we need to resize the bounding box back to the original resolution (640, 480) (width, height)
resized_bbox = tbbox.resize(fbounding_boxes, (self.width, self.height), (self.depth_img_width, self.depth_img_height))
img = timage.imresize(frame, self.depth_img_width, self.depth_img_height, 1)
# check if the bounding box is inside the 300x300 area of the GG-CNN grasping area
GGCNN_area = [190, 0, 480, 300]
GGCNN_area_center = [320, 150] # width, height
img_2 = img.asnumpy()
img = cv2.rectangle(img_2, (GGCNN_area[0], GGCNN_area[1]), (GGCNN_area[2], GGCNN_area[3]), (255, 0, 0), 1)
bbox_list, fscores_list, fclass_IDs_list = [], [], [] # bounding boxes of the chosen class
# If any object is found
if fclass_IDs.size > 0:
# If the request object is found
if self.pipeline_required_class in fclass_IDs:
print('found obj')
# we need to find all ocurrences of the class identified to consider
# situation where we have false positives as well
chosen_class_index = [i for i, x in enumerate(fclass_IDs) if x == self.pipeline_required_class]
for class_index in chosen_class_index:
bbox_list.append(resized_bbox[class_index])
fscores_list.append(fscores[class_index])
fclass_IDs_list.append(fclass_IDs[class_index])
max_score = max(fscores_list)
largest_score_bb_index = [i for i, x in enumerate(fscores_list) if x == max_score]
bbox_list = [bbox_list[largest_score_bb_index[0]]]
fscores_list = [fscores_list[largest_score_bb_index[0]]]
fclass_IDs_list = [fclass_IDs_list[largest_score_bb_index[0]]]
bbox_list = self.resize_bounding_boxes(bbox_list)
self.labels = fclass_IDs_list
self.bboxes = bbox_list
for index, bbox in enumerate(bbox_list):
# bbox_list.append(bbox)
# fscores_list.append(fscores_list[index])
# fclass_IDs_list.append(fclass_IDs_list[index])
if bbox[0] > GGCNN_area[0] and bbox[1] > GGCNN_area[1] and bbox[2] < GGCNN_area[2] and \
bbox[3] < GGCNN_area[3]:
print('obj inside ggcnn_area')
self.receive_bb_status = True
# Set the flag detection_ready
self.detection_ready.publish(True)
self.reposition_robot_flag.publish(False)
else:
print('obj outside ggcnn_area')
bbox_center_point_x = (bbox[2] - bbox[0])/2 + bbox[0] # width
bbox_center_point_y = (bbox[3] - bbox[1])/2 + bbox[1] # height
dist_x = bbox_center_point_x - GGCNN_area_center[0] # width
dist_y = GGCNN_area_center[1] - bbox_center_point_y # height
dist_x_dir = dist_x/abs(dist_x)
dist_y_dir = dist_y/abs(dist_y)
ggcnn_center_area = depth_image[GGCNN_area_center[1], GGCNN_area_center[0]]
self.horizontal_FOV = 52
self.vertical_FOV = 60
largura_2 = 2.0 * ggcnn_center_area * np.tan(self.horizontal_FOV * abs(dist_x) / depth_width_res / 2.0 / 180.0 * np.pi) / 1000 * dist_x_dir
altura_2 = 2.0 * ggcnn_center_area * np.tan(self.vertical_FOV * abs(dist_y) / depth_height_res / 2.0 / 180.0 * np.pi) / 1000 * dist_y_dir
reposition_points = Float32MultiArray()
reposition_points.data = [largura_2, altura_2]
self.reposition_coord.publish(reposition_points)
self.detection_ready.publish(True)
self.reposition_robot_flag.publish(True)
else:
print('The object ({}) was not found'.format(self.classes[self.pipeline_required_class]))
self.detection_ready.publish(False)
self.reposition_robot_flag.publish(False)
else:
print('No objects (including the requested one ({})) were found'.format(self.classes[self.pipeline_required_class]))
self.detection_ready.publish(False)
self.reposition_robot_flag.publish(False)
bbox_list = np.array(bbox_list)
fscores_list = np.array(fscores_list)
fclass_IDs_list = np.array(fclass_IDs_list)
img = gcv.utils.viz.cv_plot_bbox(img, bbox_list, fscores_list, fclass_IDs_list, class_names=self.net.classes)
depth_image = cv2.cvtColor(depth_image, cv2.COLOR_GRAY2BGR)
depth_image = depth_image.astype('uint8')
img = img.astype('uint8')
added_image = cv2.addWeighted(depth_image, 0.7, img, 0.8, 0)
self.img_pub.publish(CvBridge().cv2_to_imgmsg(added_image, 'bgr8'))