def set_car_detections(self, detections):
self.last_car_detections = detections
for det in detections:
det_label = du.find_class_detection(det)
if det_label in COLLISION_AVOIDANCE_ACTION_SET: # TODO test better distance from sign
self.actions[
ACTION_COLLISION_AVOIDANCE] = COLLISION_AVOIDANCE_DURATION_FRAMES
def find_mapping(detections, objects, object_class):
detection_object_mapping = []
for detection in detections:
# Kitty skip dontcare class
dclass = du.find_class_detection(detection)
if dclass == object_class:
max_iou = 0.0
boxA = du.find_box_detection(detection)
bestObject = None
for obj in objects:
aclass = du.find_class_gt(obj)
if aclass == object_class:
boxB = du.find_box_gt(obj)
iou = compute_IoU(boxA, boxB)
if iou > max_iou and iou > IOU_THRESHOLD: # TODO evaluate if ok set iou threshold
max_iou = iou
bestObject = obj
detection_object_mapping.append((detection, bestObject))
for obj in objects:
# Kitty skip dontcare class
aclass = du.find_class_gt(obj)
if aclass == object_class:
present = False
for det, annotation in detection_object_mapping:
if obj == annotation:
present = True
if not present:
detection_object_mapping.append((None, obj))
return detection_object_mapping
def set_sign_detections(self, detections):
with self.lock:
self.last_sign_detections = detections # only for rendering
for det in detections:
det_label = du.find_class_detection(det)
if det_label == ACTION_STOP and ACTION_STOP not in self.actions.keys(
) and ACTION_AVOID_STOP not in self.actions.keys(
): # TODO test better distance from sign
self.actions[ACTION_STOP] = STOP_DURATION_FRAMES
elif det_label == ACTION_PEDESTRIAN_CROSSING: # while sees sign and for a period slow down
self.actions[
ACTION_PEDESTRIAN_CROSSING] = PEDESTRIAN_CROSSING_DURATION_FRAMES
def compute_frame_statistics(mapping):
TP = 0
FP = 0
FN = 0
for (detection, annotation) in mapping:
if detection is None and annotation is not None: # FN
FN = FN + 1
elif detection is not None and annotation is not None: # TP
det_class = du.find_class_detection(detection)
ann_class = du.find_class_gt(annotation)
if det_class == ann_class: # Only if class is correct we have a true positive otherwise a false negative ( class always correct sinc mapping is performend only with single class)
TP = TP + 1
else:
FN = FN + 1
elif detection is not None and annotation is None: # FP
FP = FP + 1
return TP, FP, FN
def evaluate_performances(detector, test_images_path, gt_path, object_class):
images = [
f for f in listdir(test_images_path)
if isfile(join(test_images_path, f))
]
images = [f for f in images if ".png" in f]
images = [f for f in images if '._' not in f] # Avoid mac issues
final_TP = 0
final_FP = 0
final_FN = 0
cumulative_eval_time = 0.0
count = 0
for filename in images:
filenameXml = filename.replace('.png', '.xml')
img = cv2.imread(str(test_images_path) + str(filename))
im = nparray_to_image(img)
t0 = time.time()
r = detector.detect_im(im)
eval_time = time.time() - t0,
detections = detector.convert_format(r)
root = etree.parse(gt_path + filenameXml)
annotations = root.findall("object")
TP, FP, FN = evaluate_frame_performance(detections, annotations,
object_class)
final_TP = final_TP + TP
final_FP = final_FP + FP
final_FN = final_FN + FN
cumulative_eval_time = cumulative_eval_time + eval_time[0]
# TODO complete splitting performances
# ============ PRINT =============
for annotation in annotations:
aclass = du.find_class_gt(annotation)
if aclass != 'DontCare':
dr.draw_gt(img, annotation)
for detection in detections:
dclass = du.find_class_detection(detection)
if dclass != 'DontCare':
dr.draw_detection(img, detection)
# TODO remove
mapping = find_mapping(detections, annotations, object_class)
print(mapping)
for det, gt in mapping:
if det is not None:
pt1 = (int(det['topleft']['x']), int(det['topleft']['y']))
pt2 = (int(det['bottomright']['x']),
int(det['bottomright']['y']))
cv2.rectangle(img,
pt1,
pt2, (0, 0, 255),
thickness=1,
lineType=cv2.LINE_8)
if gt is not None:
box = du.find_box_gt(gt)
pt1 = (int(box[0]), int(box[1]))
pt2 = (int(box[2]), int(box[3]))
cv2.rectangle(img,
pt1,
pt2, (0, 13, 55),
thickness=1,
lineType=cv2.LINE_8)
cv2.imshow("Frame", img)
cv2.waitKey(1)
count = count + 1
print("%s %s" % (count, filename))
# =============== FINAL EVALUATION ================
precision = float(final_TP) / float(final_TP + final_FP) if float(
final_TP + final_FP) != 0.0 else np.nan
recall = float(final_TP) / float(final_TP + final_FN) if float(
final_TP + final_FN) != 0.0 else np.nan
f1_score = 2 * (precision * recall) / (precision + recall) if (
precision + recall
) != 0.0 and precision is not np.nan and recall is not np.nan else np.nan
mean_eval_time = cumulative_eval_time / len(images)
print(precision)
print(recall)
print(f1_score)
print(mean_eval_time)
return precision, recall, f1_score, mean_eval_time