CODEWITHSUNDEEP
pythonimage-processingkeras
Tien Dinh
Tien Dinh

Reputation: 361

Keras ImageDataGenerator validation_split

I am trying to train a model for facial keypoints detection from a series of frames from videos. Since these frames have almost identical position for the keypoints, the model is producing the same output for every image I put in for prediction. I was trying to use ImageDataGenerator to rotate the images in order for them to be different from each other, however, I can't seem to make it work.

Originally when I call fit on the model, I have the option to split the training data into train and validation, but I don't understand how to use the validation_split option in ImageDataGenerator. Could someone explain how to use it, or maybe suggest me a way to use this class?

Right now I have a tensor of size [total_images, width, height, channels] and its corresponding [total_images, output]. How do I use ImageDataGenerator to rotate the images, and also separate them into training and validation data?

Upvotes: 1

Views: 587

Answers (1)

Tien Dinh
Tien Dinh

Reputation: 361

Turned out I can create this myself using transformation matrices. To rotate the images correctly in OpenCV, I used this code (modified the transformation matrix to keep all the corners of the image while rotating)

Code credits: Cristian Perez Brokate you can find the explanation of the math behind this implementation. rotate_bound is exactly the same as what I found, rotate_points is the modified version of rotate_box

def rotate_bound(image, angle):
    # grab the dimensions of the image and then determine the
    # center
    (h, w) = image.shape[:2]
    (cX, cY) = (w // 2, h // 2)

    # grab the rotation matrix (applying the negative of the
    # angle to rotate clockwise), then grab the sine and cosine
    # (i.e., the rotation components of the matrix)
    M = cv2.getRotationMatrix2D((cX, cY), -angle, 1.0)
    cos = np.abs(M[0, 0])
    sin = np.abs(M[0, 1])

    # compute the new bounding dimensions of the image
    nW = int((h * sin) + (w * cos))
    nH = int((h * cos) + (w * sin))

    # adjust the rotation matrix to take into account translation
    M[0, 2] += (nW / 2) - cX
    M[1, 2] += (nH / 2) - cY

    # perform the actual rotation and return the image
    return cv2.warpAffine(image, M, (nW, nH))

And to rotate the coordinates of the points accordingly, I used this code:

def rotate_points(image, points, angle):
        # grab the dimensions of the image and then determine the
    # center
    (h, w) = image.shape[:2]
    (cX, cY) = (w // 2, h // 2)

    # grab the rotation matrix (applying the negative of the
    # angle to rotate clockwise), then grab the sine and cosine
    # (i.e., the rotation components of the matrix)
    M = cv2.getRotationMatrix2D((cX, cY), -angle, 1.0)
    cos = np.abs(M[0, 0])
    sin = np.abs(M[0, 1])

    # compute the new bounding dimensions of the image
    nW = int((h * sin) + (w * cos))
    nH = int((h * cos) + (w * sin))

    # adjust the rotation matrix to take into account translation
    M[0, 2] += (nW / 2) - cX
    M[1, 2] += (nH / 2) - cY

    v = np.ones((points.shape[0], points.shape[1] + 1))
    v[:,:-1] = points
    return np.dot(M, v.T).T

Upvotes: 1

Related Questions