def compute_ignore_margin_objectness_targets(targets: np.ndarray,
boxes: np.ndarray,
weights: np.ndarray,
pos_weights: float,
num_outputs: int) -> np.ndarray:
y_center, x_center = np_frame_ops.boxes_clipped_centers(boxes)
fm_h, fm_w = targets.shape[:2]
targets[:, :, num_outputs] = 1
for y, x, w in zip(y_center, x_center, weights):
yi = int(y * fm_h)
xi = int(x * fm_w)
targets[yi, xi, :num_outputs] = 1
targets[yi, xi, num_outputs] = pos_weights * w
for y, x, w in zip(y_center, x_center, weights):
yi = int(y * fm_h)
xi = int(x * fm_w)
for j in [yi - 1, yi, yi + 1]:
for i in [xi - 1, xi, xi + 1]:
if 0 <= j < fm_h and 0 <= i < fm_w:
if targets[j, i, 0] != 1:
targets[j, i, num_outputs] = 0
return targets
def compute_multiclass_targets(
classes_map: np.ndarray,
boxes: np.ndarray,
weights: np.ndarray,
labels: np.ndarray,
add_dustbin: bool,
dustbin_index: int,
weights_index: int,
) -> np.ndarray:
y_center, x_center = np_frame_ops.boxes_clipped_centers(boxes)
fm_height, fm_width = classes_map.shape[:2]
classes_map[:, :, weights_index] = 1
if add_dustbin:
classes_map[:, :, dustbin_index] = 1
for y, x, w, label_index in zip(y_center, x_center, weights, labels):
yi = int(y * fm_height)
xi = int(x * fm_width)
classes_map[yi, xi, label_index] = 1.0
classes_map[yi, xi, weights_index] = w
if add_dustbin:
classes_map[yi, xi, dustbin_index] = 0.0
return classes_map
def compute_box_offset_targets(offset_map: np.ndarray, boxes: np.ndarray) -> np.ndarray:
y_center, x_center = np_frame_ops.boxes_clipped_centers(boxes)
fm_height, fm_width = offset_map.shape[:2]
for k, (y, x) in enumerate(zip(y_center, x_center)):
yi = int(y * fm_height)
xi = int(x * fm_width)
offset_map[yi, xi, 0] = y * fm_height - yi
offset_map[yi, xi, 1] = x * fm_width - xi
offset_map[yi, xi, 2] = 1
return offset_map
def compute_box_size_targets(hw_map: np.ndarray, boxes: np.ndarray) -> np.ndarray:
y_center, x_center = np_frame_ops.boxes_clipped_centers(boxes)
heights, widths = np_frame_ops.boxes_heights_widths(boxes)
fm_height, fm_width = hw_map.shape[:2]
for k, (y, x) in enumerate(zip(y_center, x_center)):
yi = int(y * fm_height)
xi = int(x * fm_width)
hw_map[yi, xi, 0] = heights[k] * fm_height
hw_map[yi, xi, 1] = widths[k] * fm_width
hw_map[yi, xi, 2] = 1
return hw_map
def compute_objectness_targets(targets: np.ndarray, boxes: np.ndarray,
weights: np.ndarray, pos_weights: float,
num_outputs: int) -> np.ndarray:
y_center, x_center = np_frame_ops.boxes_clipped_centers(boxes)
fm_height, fm_width = targets.shape[:2]
targets[:, :, num_outputs] = 1
for y, x, w in zip(y_center, x_center, weights):
yi = int(y * fm_height)
xi = int(x * fm_width)
targets[yi, xi, :num_outputs] = 1
targets[yi, xi, num_outputs] = pos_weights * w
return targets
def compute_mean_box_size_targets(hw_map: np.ndarray, boxes: np.ndarray) -> np.ndarray:
y_center, x_center = np_frame_ops.boxes_clipped_centers(boxes)
heights, widths = np_frame_ops.boxes_heights_widths(boxes)
fm_height, fm_width = hw_map.shape[:2]
for k, (y, x) in enumerate(zip(y_center, x_center)):
yi = int(y * fm_height)
xi = int(x * fm_width)
hw_map[yi, xi, 0] += heights[k] * fm_height
hw_map[yi, xi, 1] += widths[k] * fm_width
hw_map[yi, xi, 2] += 1
hw_map[:, :, :2] = hw_map[:, :, :2] / (hw_map[:, :, 2:] + 1e-6)
# make weights in range (0, 1)
hw_map[:, :, 2] = np.minimum(hw_map[:, :, 2], 1)
return hw_map
def compute_smooth_objectness_targets(targets: np.ndarray, boxes: np.ndarray,
weights: np.ndarray, pos_weights: float,
num_outputs: int) -> np.ndarray:
y_center, x_center = np_frame_ops.boxes_clipped_centers(boxes)
fm_h, fm_w = targets.shape[:2]
targets[:, :, num_outputs] = 1
for y, x, w in zip(y_center, x_center, weights):
yi = int(y * fm_h)
xi = int(x * fm_w)
for j in [yi - 1, yi, yi + 1]:
for i in [xi - 1, xi, xi + 1]:
y_offset = y * fm_h - j
x_offset = x * fm_w - i
if 0 <= j < fm_h and 0 <= i < fm_w:
delta = -0.5 * (y_offset**2 + x_offset**2)
targets[j, i, :num_outputs] = np.exp(delta)
targets[j, i, num_outputs] = pos_weights * w
return targets