Detect objects in an image with Haar Cascade Classifier.
The HaarCascadeClassifier, when used to detect objects, is returning a list of detections, with weights instead of confidence scores. This class converts weights into scores ranging [0.998, min_ratio] with a modified version of the Unity-based normalization method.
This classifier already delete overlapped detections and allows to fix a threshold size to filter too small objects.
def __init__(self):
super(HaarCascadeDetector, self).__init__()
self._extension = '.xml'
Return the normalized list of values.
Use the Unity-based normalization, slightly adapted.
def normalize_weights(self, dataset):
"""Return the normalized list of values.
Use the Unity-based normalization, slightly adapted.
:param dataset: (list) List of float weight values
:returns: list of confidence scores
"""
a = self.get_min_score()
b = 0.998
coeff = b - a
norm_list = list()
if isinstance(dataset, list):
min_value = min(dataset) * b
max_value = max(dataset) * 1.01
for value in dataset:
tmp = a + (value - min_value) * coeff / (max_value - min_value)
norm_list.append(min(1.0, max(0.0, tmp)))
return norm_list
Initialize the detector with the given file.
Exception
def _set_detector(self, model):
"""Initialize the detector with the given file.
:param model: (str) Filename of the XML Haar Cascade file
:raises: Exception
"""
try:
self._detector = cv2.CascadeClassifier(model)
except cv2.error as e:
logging.error('Loading the XML Haar Cascade model failed.')
raise sppasError(str(e))
Determine the coordinates of the detected objects.
def _detection(self, image):
"""Determine the coordinates of the detected objects.
:param image: (sppasImage or numpy.ndarray)
"""
detections = self.__haar_detections(image)
if len(detections[0]) == 0:
return
scores = list()
for d in detections[2]:
if isinstance(d, (list, tuple)) is True:
scores.append(d[0])
else:
scores.append(d)
normalized = self.normalize_weights(scores)
for rect, score in zip(detections[0], normalized):
coords = sppasCoords(rect[0], rect[1], rect[2], rect[3])
coords.set_confidence(score)
self._coords.append(coords)
Detect objects using the Haar Cascade classifier.
This classifier already delete overlapped detections and too small detected objects. Returned detections are a tuple with 3 values: - a list of N-list of rectangles; - a list of N times the same int value (why???); - a list of N weight values.
Notice that the scale factor has a big impact on the estimation time: - with 1.04 => 5.3x real time, - with 1.10 => 2.2x real time, but -6% of detected objects (relevant ones?)
def __haar_detections(self, image, scale_factor=1.06):
"""Detect objects using the Haar Cascade classifier.
This classifier already delete overlapped detections and too small
detected objects. Returned detections are a tuple with 3 values:
- a list of N-list of rectangles;
- a list of N times the same int value (why???);
- a list of N weight values.
Notice that the scale factor has a big impact on the estimation time:
- with 1.04 => 5.3x real time,
- with 1.10 => 2.2x real time, but -6% of detected objects (relevant ones?)
:param image: (sppasImage)
:param scale_factor: (float) how much the image size is reduced at each image scale
:return: (numpy arrays)
"""
w, h = image.size()
min_w = int(float(w) * self.get_min_ratio())
min_h = int(float(h) * self.get_min_ratio())
try:
detections = self._detector.detectMultiScale3(image, scaleFactor=scale_factor, minNeighbors=3, minSize=(min_w, min_h), flags=0, outputRejectLevels=True)
except cv2.error as e:
self._coords = list()
raise sppasError('HaarCascadeClassifier detection failed: {}'.format(str(e)))
return detections