def accumulate(self, scoremap, image_id):
"""
Score histograms over the score map values at GT positive and negative
pixels are computed.
Args:
scoremap: numpy.ndarray(size=(H, W), dtype=np.float)
image_id: string.
"""
check_scoremap_validity(scoremap)
gt_mask = get_mask(self.mask_root,
self.mask_paths[image_id],
self.ignore_paths[image_id])
gt_true_scores = scoremap[gt_mask == 1]
gt_false_scores = scoremap[gt_mask == 0]
# histograms in ascending order
gt_true_hist, _ = np.histogram(gt_true_scores,
bins=self.threshold_list_right_edge)
self.gt_true_score_hist += gt_true_hist.astype(np.float)
gt_false_hist, _ = np.histogram(gt_false_scores,
bins=self.threshold_list_right_edge)
self.gt_false_score_hist += gt_false_hist.astype(np.float)
def compute_bboxes_from_scoremaps(scoremap,
scoremap_threshold_list,
multi_contour_eval=False):
"""
Args:
scoremap: numpy.ndarray(dtype=np.float32, size=(H, W)) between 0 and 1
scoremap_threshold_list: iterable
multi_contour_eval: flag for multi-contour evaluation
Returns:
estimated_boxes_at_each_thr: list of estimated boxes (list of np.array)
at each cam threshold
number_of_box_list: list of the number of boxes at each cam threshold
"""
check_scoremap_validity(scoremap)
height, width = scoremap.shape
scoremap_image = np.expand_dims((scoremap * 255).astype(np.uint8), 2)
def scoremap2bbox(threshold):
_, thr_gray_heatmap = cv2.threshold(src=scoremap_image,
thresh=int(threshold *
np.max(scoremap_image)),
maxval=255,
type=cv2.THRESH_BINARY)
contours = cv2.findContours(
image=thr_gray_heatmap,
mode=cv2.RETR_TREE,
method=cv2.CHAIN_APPROX_SIMPLE)[_CONTOUR_INDEX]
if len(contours) == 0:
return np.asarray([[0, 0, 0, 0]]), 1
if not multi_contour_eval:
contours = [max(contours, key=cv2.contourArea)]
estimated_boxes = []
for contour in contours:
x, y, w, h = cv2.boundingRect(contour)
x0, y0, x1, y1 = x, y, x + w, y + h
x1 = min(x1, width - 1)
y1 = min(y1, height - 1)
estimated_boxes.append([x0, y0, x1, y1])
return np.asarray(estimated_boxes), len(contours)
estimated_boxes_at_each_thr = []
number_of_box_list = []
for threshold in scoremap_threshold_list:
boxes, number_of_box = scoremap2bbox(threshold)
estimated_boxes_at_each_thr.append(boxes)
number_of_box_list.append(number_of_box)
return estimated_boxes_at_each_thr, number_of_box_list
def compute_bboxes_from_scoremaps(scoremap, scoremap_threshold_list):
"""
Args:
scoremap: numpy.ndarray(dtype=np.float32, size=(H, W)) between 0 and 1
scoremap_threshold_list: iterable
Returns:
boxes: list of estimated boxes (list of ints) at each cam threshold
"""
check_scoremap_validity(scoremap)
height, width = scoremap.shape
scoremap_image = np.expand_dims((scoremap * 255).astype(np.uint8), 2)
def scoremap2bbox(threshold):
_, thr_gray_heatmap = cv2.threshold(src=scoremap_image,
thresh=int(threshold *
np.max(scoremap_image)),
maxval=255,
type=cv2.THRESH_BINARY)
contours = cv2.findContours(image=thr_gray_heatmap,
mode=cv2.RETR_TREE,
method=cv2.CHAIN_APPROX_SIMPLE)[1]
if len(contours) == 0:
return [0, 0, 0, 0]
c = max(contours, key=cv2.contourArea)
x, y, w, h = cv2.boundingRect(c)
x0, y0, x1, y1 = x, y, x + w, y + h
x1 = min(x1, width - 1)
y1 = min(y1, height - 1)
return [x0, y0, x1, y1]
estimated_bbox = [
scoremap2bbox(threshold) for threshold in scoremap_threshold_list
]
return estimated_bbox