p.x = p.x * width + t_x
p.y = p.y * height + t_y
Results in:
My code is below. I would be happy if you could have a look. In the comments there is a method that gives the correct result visually, but is not correct as it does not transform the x, y coordinates but just shifts the resulting bounding box. The idea is to use translation as a data augmentation technique, which augments both input images and corresponding x, y annotations so that the neural network does not simpy memorize the input data and the corresponding annotations.
def random_translation(image_path, bounding_box, factor = 20):
colors = [[255, 0, 0], [0, 255, 0], [0, 0, 255],
[255, 255, 0], [255, 127, 127], [255, 165, 0],
[255, 105, 180], [64, 224, 208], [134, 1, 175],
[127, 127, 127], [116, 86, 74], [0, 0, 0],
[128, 0, 128], [0, 128, 128], [128, 0, 0],
[0, 255, 255], [128, 128, 128], [255, 0, 255],
[0, 0, 128], [255, 105, 180], [128, 128, 0]]
class_names = ["aeroplane","bicycle","bird","boat","bottle","bus","car",
"cat","chair","cow","diningtable","dog","horse","motorbike","person",
"pottedplant","sheep","sofa","train","tvmonitor"]
image = cv.imread(image_path)
height, width = image.shape[:2]
# 0. Translation Variables
translate_upper_bound = float(height / 100 * factor)
translate_lower_bound = float(height / 100 * factor) * -1
# uniform vals to translate into x coord t_x and y coord t_y
t_x = np.random.uniform(low=translate_lower_bound, high=translate_upper_bound)
t_y = np.random.uniform(low=translate_lower_bound, high=translate_upper_bound)
# Translation matrix T
T = np.float32([[1, 0, t_x], [0, 1, t_y]])
img_translation = cv.warpAffine(image, T, (width, height))
# 1. Original Image and Bounding Box
height, width, _ = image.shape
class_pred = int(bounding_box[0])
bounding_box = bounding_box[1:]
assert len(bounding_box) == 4, "Bounding box prediction exceed x,y,w,h."
# extract x midpoint, y midpoint, w width and h height
x = bounding_box[0]
y = bounding_box[1]
w = bounding_box[2]
h = bounding_box[3]
l = int((x - w / 2) * width)
r = int((x + w / 2) * width)
t = int((y - h / 2) * height)
b = int((y + h / 2) * height)
if l < 0:
l = 0
if r > width - 1:
r = width - 1
if t < 0:
t = 0
if b > height - 1:
b = height - 1
image = cv.rectangle(image, (l, t), (r, b), colors[class_pred], 2)
(width, height), _ = cv.getTextSize(class_names[class_pred], cv.FONT_HERSHEY_SIMPLEX, 0.6, 2)
image = cv.rectangle(image, (l, t + 20), (l + width, t), colors[class_pred], -1)
image = cv.putText(image, class_names[class_pred], (l, t + 15),
cv.FONT_HERSHEY_SIMPLEX, 0.6, [255, 255, 255], 2)
#2. Translated Image and Bounding Box
height, width, _ = img_translation.shape
x = bounding_box[0] * width + t_x
y = bounding_box[1] * height + t_y
w = bounding_box[2]
h = bounding_box[3]
# this works but is incorrect, as it doesnt transform the annotated x, y labels
# it draws the box where it would be and then shifts the entire bounding box
#l = int((x - w / 2) * width + t_x)
#r = int((x + w / 2) * width + t_x)
#t = int((y - h / 2) * height + t_y)
#b = int((y + h / 2) * height + t_y)
l = int((x - w / 2) * width)
r = int((x + w / 2) * width)
t = int((y - h / 2) * height)
b = int((y + h / 2) * height)
if l < 0:
l = 0
if r > width - 1:
r = width - 1
if t < 0:
t = 0
if b > height - 1:
b = height - 1
img_translation = cv.rectangle(img_translation, (l, t), (r, b), colors[class_pred], 2)
(width, height), _ = cv.getTextSize(class_names[class_pred], cv.FONT_HERSHEY_SIMPLEX, 0.6, 2)
img_translation = cv.rectangle(img_translation, (l, t + 20), (l + width, t), colors[class_pred], -1)
img_translation = cv.putText(img_translation, class_names[class_pred], (l, t + 15),
cv.FONT_HERSHEY_SIMPLEX, 0.6, [255, 255, 255], 2)
# 3. Plot Results
plt.subplot(1,2,1)
plt.imshow(image)
plt.subplot(1,2,2)
plt.imshow(img_translation)
bbox = [11, 0.34419263456090654, 0.611, 0.4164305949008499, 0.262]
random_translation(image_path = 'C:/Users/username/anaconda3/envs/yolo/yolo/data/images/000001.jpg', bounding_box = bbox, factor= 20)
