def _load_crop_helper(self, img_file_name, img_bbox):
img_whole_im = load_image_tensorflow(img_file_name, jpg=self.jpg, channels=3)
dims = tf.shape(img_whole_im)
img_x = img_bbox[0]
img_y = img_bbox[1]
img_w = img_bbox[2]
img_h = img_bbox[3]
# add context region
img_x -= 0.5 * img_w * (self.context_region_factor - 1.0)
img_y -= 0.5 * img_h * (self.context_region_factor - 1.0)
img_w *= self.context_region_factor
img_h *= self.context_region_factor
# round to integer coordinates
img_x = tf.cast(tf.round(img_x), tf.int32)
img_y = tf.cast(tf.round(img_y), tf.int32)
img_w = tf.cast(tf.round(img_w), tf.int32)
img_h = tf.cast(tf.round(img_h), tf.int32)
# clip to image size
img_x = tf.maximum(img_x, 0)
img_y = tf.maximum(img_y, 0)
img_excess_w = tf.maximum(img_x + img_w - dims[1], 0)
img_excess_h = tf.maximum(img_y + img_h - dims[0], 0)
img_w = img_w - img_excess_w
img_h = img_h - img_excess_h
# crop
img_cropped = img_whole_im[img_y:img_y + img_h, img_x:img_x + img_w]
# resize
img = resize_image(img_cropped, self.input_size, True)
img.set_shape(self.input_size + (3,))
# augment and normalize
tensors = {"unnormalized_img": img}
tensors = apply_augmentors(tensors, self.augmentors)
img = tensors["unnormalized_img"]
img_norm = normalize(img)
return img_norm, img, img_cropped
def _load_crop_helper(self, img_file_name, img_bbox):
img = load_image_tensorflow(img_file_name, jpg=self.jpg, channels=3)
img.set_shape(self.input_size + (3, ))
# augment and normalize
tensors = {"unnormalized_img": img}
tensors = apply_augmentors(tensors, self.augmentors)
img = tensors["unnormalized_img"]
img_norm = normalize(img)
return img_norm, img, None
def do_one_crop(self, box):
# Do the crop and resize myself
x = tf.cast(box[0], tf.int32)
y = tf.cast(box[1], tf.int32)
w = tf.cast(box[2], tf.int32)
h = tf.cast(box[3], tf.int32)
# image = tf.expand_dims(self.image, 0)
img_cropped = self.conv_image[y:y + h, x:x + w]
min_dim = tf.minimum(h, w)
# # resize
# img = resize_image(img_cropped, self.input_size, True)
# resize
img = tf.cond(
min_dim > 10,
lambda: resize_image(img_cropped, self.input_size, True),
lambda: tf.zeros([self.input_size[0], self.input_size[1], 3]))
# img.set_shape(self.input_size + (3,))
norm_img = normalize(img)
return norm_img
def load_normalized_image_tensorflow(im_path, jpg):
img = load_image_tensorflow(im_path, jpg=jpg)
img = normalize(img)
return img
def assemble_input_tensors(tensors, img_size=(None, None)):
img_size = list(img_size)
assert "unnormalized_img" in tensors
assert "tag" in tensors
img = tensors["unnormalized_img"]
img = normalize(img)
tag = tensors["tag"]
if "label" in tensors:
label = tensors["label"]
else:
label = None
if "raw_label" in tensors:
raw_label = tensors["raw_label"]
else:
raw_label = label
concats = [img]
n_input_channels = 3
if "old_label" in tensors:
old_label = tf.cast(tensors["old_label"], tf.float32)
if Constants.OLD_LABEL_AS_DT in tensors and tensors[Constants.OLD_LABEL_AS_DT] is True:
clip_value = tf.ones_like(old_label) * 255
old_label = tf.where(tf.greater(old_label, 255), clip_value, old_label)
old_label /= 255.0
concats.append(old_label)
n_input_channels += 1
if Constants.DT_NEG in tensors:
# Do not use the click channel as they can be deciphered from the distance transforms.
u0 = tensors[Constants.DT_NEG]
u0 = tf.cast(u0, tf.float32)
clip_value = tf.ones_like(u0)*255
u0 = tf.where(tf.greater(u0, 255), clip_value, u0)
u0 /= 255.0
concats.append(u0)
n_input_channels += u0.get_shape().as_list()[-1]
if Constants.DT_POS in tensors:
u1 = tensors[Constants.DT_POS]
u1 = tf.cast(u1, tf.float32)
clip_value = tf.ones_like(u1) * 255
u1 = tf.where(tf.greater(u1, 255), clip_value, u1 )
u1 /= 255.0
concats.append(u1)
n_input_channels += u1.get_shape().as_list()[-1]
if "flow_past" in tensors:
concats.append(tensors["flow_past"])
n_input_channels += 2
if "flow_future" in tensors:
concats.append(tensors["flow_future"])
n_input_channels += 2
if len(concats) > 1:
img = tf.concat(concats, axis=2)
img.set_shape(img_size + [n_input_channels])
tensors_out = {"inputs": img, "tags": tag}
if label is not None:
tensors_out["labels"] = label
if raw_label is not None:
tensors_out["raw_labels"] = raw_label
if "index_img" in tensors:
tensors_out["index_imgs"] = tensors["index_img"]
#TODO: maybe just start with a copy of tensors instead of doing every case separately?
if Constants.BBOXES in tensors:
tensors_out[Constants.BBOXES] = tensors[Constants.BBOXES]
if Constants.IDS in tensors:
tensors_out[Constants.IDS] = tensors[Constants.IDS]
if Constants.CLASSES in tensors:
tensors_out[Constants.CLASSES] = tensors[Constants.CLASSES]
if Constants.IMG_IDS in tensors:
tensors_out[Constants.IMG_IDS] = tensors[Constants.IMG_IDS]
if Constants.ORIGINAL_SIZES in tensors:
tensors_out[Constants.ORIGINAL_SIZES] = tensors[Constants.ORIGINAL_SIZES]
if Constants.RESIZED_SIZES in tensors:
tensors_out[Constants.RESIZED_SIZES] = tensors[Constants.RESIZED_SIZES]
if Constants.IGNORE_REGIONS in tensors:
tensors_out[Constants.IGNORE_REGIONS] = tensors[Constants.IGNORE_REGIONS]
if Constants.SCENE_INFOS in tensors:
tensors_out[Constants.SCENE_INFOS] = tensors[Constants.SCENE_INFOS]
if Constants.OLD_LABEL_AS_DT in tensors:
tensors_out[Constants.OLD_LABEL_AS_DT] = tensors[Constants.OLD_LABEL_AS_DT]
return tensors_out