def _predict(self, img_msg, label_msg):
# convert image
bridge = cv_bridge.CvBridge()
input_image = bridge.imgmsg_to_cv2(img_msg, 'rgb8')
input_label = bridge.imgmsg_to_cv2(label_msg)
# predict
region_imgs = []
for l in np.unique(input_label):
if l == 0: # bg_label
continue
mask = (input_label == l)
region = jsk_recognition_utils.bounding_rect_of_mask(
input_image, mask)
region_imgs.append(region)
y_proba = self.estimator.predict(region_imgs)
target_names = np.array(jsk_apc2015_common.get_object_list())
y_pred = np.argmax(y_proba, axis=-1)
label_proba = [p[i] for p, i in zip(y_proba, y_pred)]
# prepare message
res = ClassificationResult()
res.header = img_msg.header
res.labels = y_pred
res.label_names = target_names[y_pred]
res.label_proba = label_proba
res.probabilities = y_proba.reshape(-1)
res.classifier = '<jsk_2015_05_baxter_apc.ColorHistogramFeatures>'
res.target_names = target_names
self._pub.publish(res)
def _apply(self, bof_msg, ch_msg):
target_names = bof_msg.target_names
assert target_names == ch_msg.target_names
N_label = len(target_names)
bof_proba = np.array(bof_msg.probabilities).reshape((-1, N_label))
ch_proba = np.array(ch_msg.probabilities).reshape((-1, N_label))
bof_weight = np.array([self.weight[n]['bof'] for n in target_names])
ch_weight = np.array([self.weight[n]['color'] for n in target_names])
y_proba = (bof_weight * bof_proba) + (ch_weight * ch_proba)
# verification result for debug
y_pred = np.argmax(y_proba, axis=-1)
target_names = np.array(target_names)
label_proba = [p[i] for p, i in zip(y_proba, y_pred)]
# compose msg
msg = ClassificationResult()
msg.header = bof_msg.header
msg.labels = y_pred
msg.label_names = target_names[y_pred]
msg.label_proba = label_proba
msg.probabilities = y_proba.reshape(-1)
msg.classifier = '<jsk_2015_05_baxter_apc.BoostObjectRecognition>'
msg.target_names = target_names
self.pub.publish(msg)
def _apply(self, bof_msg, ch_msg):
target_names = bof_msg.target_names
assert target_names == ch_msg.target_names
N_label = len(target_names)
bof_proba = np.array(bof_msg.probabilities).reshape((-1, N_label))
ch_proba = np.array(ch_msg.probabilities).reshape((-1, N_label))
bof_weight = np.array([self.weight[n]['bof'] for n in target_names])
ch_weight = np.array([self.weight[n]['color'] for n in target_names])
y_proba = (bof_weight * bof_proba) + (ch_weight * ch_proba)
# verification result for debug
y_pred = np.argmax(y_proba, axis=-1)
target_names = np.array(target_names)
label_proba = [p[i] for p, i in zip(y_proba, y_pred)]
# compose msg
msg = ClassificationResult()
msg.header = bof_msg.header
msg.labels = y_pred
msg.label_names = target_names[y_pred]
msg.label_proba = label_proba
msg.probabilities = y_proba.reshape(-1)
msg.classifier = '<jsk_2015_05_baxter_apc.BoostObjectRecognition>'
msg.target_names = target_names
self.pub.publish(msg)
def _cb(self, imgmsg):
bridge = cv_bridge.CvBridge()
img = bridge.imgmsg_to_cv2(imgmsg)
label, proba = self._classify(img)
msg = ClassificationResult()
msg.header = imgmsg.header
msg.labels = [label]
msg.label_names = [self.target_names[label]]
msg.label_proba = [proba[label]]
msg.probabilities = proba
msg.classifier = self.classifier_name
msg.target_names = self.target_names
self.pub.publish(msg)
def _cb(self, imgmsg):
bridge = cv_bridge.CvBridge()
img = bridge.imgmsg_to_cv2(imgmsg)
label, proba = self._classify(img)
msg = ClassificationResult()
msg.header = imgmsg.header
msg.labels = [label]
msg.label_names = [self.target_names[label]]
msg.label_proba = [proba[label]]
msg.probabilities = proba
msg.classifier = self.classifier_name
msg.target_names = self.target_names
self.pub.publish(msg)
def callback(self, imgmsg, mask_msg):
bridge = cv_bridge.CvBridge()
img = bridge.imgmsg_to_cv2(imgmsg, desired_encoding='bgr8')
mask = bridge.imgmsg_to_cv2(mask_msg, desired_encoding='mono8')
if mask.ndim == 3:
mask = np.squeeze(mask, axis=2)
mask = mask >= 127 # uint8 -> bool
img = img.astype(np.float64)
img[mask] -= self.mean[mask]
img[~mask] = 0
img = img.transpose(2, 0, 1)
img = img.astype(np.float32)
x_data = np.asarray([img])
if self.gpu >= 0:
x_data = cuda.to_gpu(x_data)
x = chainer.Variable(x_data)
y = self.model(x)
feat = cuda.to_cpu(y.data)
feat = feat.squeeze(axis=(2, 3))
X_query = feat
y_pred_proba = self.knn.predict_proba(X_query)
y_pred = np.argmax(y_pred_proba, axis=1)
classes = self.knn.classes_
target_names = self.target_names[classes]
msg = ClassificationResult()
msg.header = imgmsg.header
msg.labels = y_pred.tolist()
msg.label_names = target_names[y_pred].tolist()
msg.label_proba = y_pred_proba[:, y_pred].flatten().tolist()
msg.probabilities = y_pred_proba.flatten().tolist()
msg.target_names = target_names.tolist()
self.pub.publish(msg)
def callback(self, imgmsg, mask_msg):
bridge = cv_bridge.CvBridge()
img = bridge.imgmsg_to_cv2(imgmsg, desired_encoding='bgr8')
mask = bridge.imgmsg_to_cv2(mask_msg, desired_encoding='mono8')
if mask.ndim == 3:
mask = np.squeeze(mask, axis=2)
mask = mask >= 127 # uint8 -> bool
img = img.astype(np.float64)
img[mask] -= self.mean[mask]
img[~mask] = 0
img = img.transpose(2, 0, 1)
img = img.astype(np.float32)
x_data = np.asarray([img])
if self.gpu >= 0:
x_data = cuda.to_gpu(x_data)
x = chainer.Variable(x_data)
y = self.model(x)
feat = cuda.to_cpu(y.data)
feat = feat.squeeze(axis=(2, 3))
X_query = feat
y_pred_proba = self.knn.predict_proba(X_query)
y_pred = np.argmax(y_pred_proba, axis=1)
classes = self.knn.classes_
target_names = self.target_names[classes]
msg = ClassificationResult()
msg.header = imgmsg.header
msg.labels = y_pred.tolist()
msg.label_names = target_names[y_pred].tolist()
msg.label_proba = y_pred_proba[:, y_pred].flatten().tolist()
msg.probabilities = y_pred_proba.flatten().tolist()
msg.target_names = target_names.tolist()
self.pub.publish(msg)