def visualize_classification(image, post_processed, config):
"""Draw classfication result to inference input image.
Args:
image (np.ndarray): A inference input RGB image to be draw.
post_processed (np.ndarray): A one batch output of model be already applied post process.
format is defined at https://github.com/blue-oil/blueoil/blob/master/lmnet/docs/specification/output_data.md
config (EasyDict): Inference config.
Returns:
PIL.Image.Image: drawn image object.
"""
colormap = get_color_map(len(config.CLASSES))
font_size = 16
max_class_id = int(np.argmax(post_processed))
image_width = image.shape[1]
image = PIL.Image.fromarray(image)
draw = PIL.ImageDraw.Draw(image)
font = PIL.ImageFont.truetype(FONT, font_size)
text_color = colormap[max_class_id]
class_name = config.CLASSES[max_class_id]
score = float(np.max(post_processed))
text = "{}\n{:.3f}".format(class_name, score)
text_size = draw.multiline_textsize(text, font=font)
draw.multiline_text((image_width - text_size[0] - 10, 0), text, fill=text_color, font=font, align="right")
return image
def __init__(self, *args, label_colors=None, **kwargs):
super().__init__(
*args,
**kwargs,
)
if label_colors is None:
self.label_colors = get_color_map(self.num_classes)
else:
self.label_colors = label_colors
def run_sementic_segmentation(config):
global nn
camera_width = 320
camera_height = 240
window_name = "Segmentation Demo"
vc = init_camera(camera_width, camera_height)
pool = Pool(processes=1, initializer=nn.init)
result = None
fps = 1.0
q_save = Queue()
q_show = Queue()
grabbed, camera_img = vc.read()
if not grabbed:
print("Frame is empty")
q_show.put(camera_img.copy())
input_img = camera_img.copy()
colormap = np.array(get_color_map(len(config['CLASSES'])), dtype=np.uint8)
while True:
m1 = MyTime("1 loop of while(1) of main()")
pool_result = pool.apply_async(_run_inference, (input_img, ))
is_first = True
while True:
grabbed, camera_img = vc.read()
if is_first:
input_img = camera_img.copy()
is_first = False
q_save.put(camera_img.copy())
if not q_show.empty():
window_img = q_show.get()
overlay_img = window_img
if result is not None:
seg_img = label_to_color_image(result, colormap)
seg_img = cv2.resize(seg_img,
dsize=(camera_width, camera_height))
overlay_img = cv2.addWeighted(window_img, 1, seg_img, 0.8,
0)
overlay_img = add_fps(overlay_img, fps)
cv2.imshow(window_name, overlay_img)
key = cv2.waitKey(2) # Wait for 2ms
if key == 27: # ESC to quit
return
if pool_result.ready():
break
q_show = clear_queue(q_show)
q_save, q_show = swap_queue(q_save, q_show)
result, fps = pool_result.get()
m1.show()
def visualize_object_detection(image, post_processed, config):
"""Draw object detection result boxes to image.
Args:
image (np.ndarray): A inference input RGB image to be draw.
post_processed (np.ndarray): A one batch output of model be already applied post process.
format is defined at https://github.com/blue-oil/blueoil/blob/master/lmnet/docs/specification/output_data.md
config (EasyDict): Inference config.
Returns:
PIL.Image.Image: drawn image object.
"""
colormap = get_color_map(len(config.CLASSES))
height_scale = image.shape[0] / float(config.IMAGE_SIZE[0])
width_scale = image.shape[1] / float(config.IMAGE_SIZE[1])
predict_boxes = np.copy(post_processed)
predict_boxes[:, 0] *= width_scale
predict_boxes[:, 1] *= height_scale
predict_boxes[:, 2] *= width_scale
predict_boxes[:, 3] *= height_scale
image = PIL.Image.fromarray(image)
draw = PIL.ImageDraw.Draw(image)
for predict_box in predict_boxes:
class_id = int(predict_box[4])
class_name = config.CLASSES[class_id]
box = [x for x in predict_box[:4]]
score = predict_box[5]
color = tuple(colormap[class_id])
xy = [box[0], box[1], box[0] + box[2], box[1] + box[3]]
draw.rectangle(xy, outline=color)
top_left = [box[0] - 10, box[1]]
txt = "{:s}: {:.3f}".format(class_name, float(score))
draw.text(top_left, txt, fill=color)
return image
def visualize_semantic_segmentation(image, post_processed, config):
"""Draw semantic segmentation result mask to image.
Args:
image (np.ndarray): A inference input RGB image to be draw.
post_processed (np.ndarray): A one batch output of model be already applied post process.
format is defined at https://github.com/blue-oil/blueoil/blob/master/lmnet/docs/specification/output_data.md
config (EasyDict): Inference config.
Returns:
iamge (PIL.Image.Image): drawn image object.
"""
colormap = np.array(get_color_map(len(config.CLASSES)), dtype=np.uint8)
alpha = 0.5
image_height = image.shape[0]
image_width = image.shape[1]
mask_image = label_to_color_image(np.expand_dims(post_processed, 0), colormap)
mask_img = PIL.Image.fromarray(mask_image)
mask_img = mask_img.resize(size=(image_width, image_height))
result = PIL.Image.blend(PIL.Image.fromarray(image), mask_img, alpha)
return result
def __init__(self, task, classes, image_size):
assert task in Tasks
self.task = task
self.classes = classes
self.image_size = image_size
self.color_maps = get_color_map(len(classes))
def test_get_color_map_with_large_length():
color_map = get_color_map(30)
assert len(color_map) == 30
assert color_map[0] == (192, 0, 128)
assert color_map[29] == (128, 0, 192)
def test_get_color_map_with_small_length():
color_map = get_color_map(5)
assert len(color_map) == 5
assert color_map[0] == (192, 0, 128)
assert color_map[4] == (64, 0, 128)
def label_colors(self):
classes = self.parse_labels()
return get_color_map(len(classes))