Function generatepatchs

src/depthmap_generation.py:1070–1099

(img, base_size, factor)

# Generating local patches to perform the local refinement described in section 6 of the main paper.
def generatepatchs(img, base_size, factor):
    # Compute the gradients as a proxy of the contextual cues.
    img_gray = rgb2gray(img)
    whole_grad = np.abs(cv2.Sobel(img_gray, cv2.CV_64F, 0, 1, ksize=3)) + \
                 np.abs(cv2.Sobel(img_gray, cv2.CV_64F, 1, 0, ksize=3))

    threshold = whole_grad[whole_grad > 0].mean()
    whole_grad[whole_grad < threshold] = 0

    # We use the integral image to speed-up the evaluation of the amount of gradients for each patch.
    gf = whole_grad.sum() / len(whole_grad.reshape(-1))
    grad_integral_image = cv2.integral(whole_grad)

    # Variables are selected such that the initial patch size would be the receptive field size
    # and the stride is set to 1/3 of the receptive field size.
    blsize = int(round(base_size / 2))
    stride = int(round(blsize * 0.75))

    # Get initial Grid
    patch_bound_list = applyGridpatch(blsize, stride, img, [0, 0, 0, 0])

    # Refine initial Grid of patches by discarding the flat (in terms of gradients of the rgb image) ones. Refine
    # each patch size to ensure that there will be enough depth cues for the network to generate a consistent depth map.
    print("Selecting patches ...")
    patch_bound_list = adaptiveselection(grad_integral_image, patch_bound_list, gf, factor)

    # Sort the patch list to make sure the merging operation will be done with the correct order: starting from biggest
    # patch
    patchset = sorted(patch_bound_list.items(), key=lambda x: getitem(x[1], 'size'), reverse=True)
    return patchset


def applyGridpatch(blsize, stride, img, box):