SAD(Sum of Absolute Differences)
def matte_sad(pred_matte, gt_matte):
'''
Sum of Absolute Differences
pred_matte : np.array, shape : [h,w]
gt_matte : np.array, shape : [h,w]
'''
assert(len(pred_matte.shape) == len(gt_matte.shape))
error_sad = np.sum(np.abs(pred_matte - gt_matte))
return error_sad
MSE(Mean Squared Error)
def matte_mse(pred_matte, gt_matte):
''' Mean Squared Error '''
assert(len(pred_matte.shape) == len(gt_matte.shape))
error_mse = np.mean(np.power(pred_matte-gt_matte, 2))
return error_mse
Gradient
def matte_grad(pred_matte, gt_matte):
''' Error measure with Gradient '''
assert(len(pred_matte.shape) == len(gt_matte.shape))
predict_grad = scipy.ndimage.filters.gaussian_filter(pred_matte, 1.4, order=1)
gt_grad = scipy.ndimage.filters.gaussian_filter(gt_matte, 1.4, order=1)
error_grad = np.sum(np.power(predict_grad - gt_grad, 2))
return error_grad
Connectivity
1
参考文献：
[1]. Christoph Rhemann, et al. A Perceptually Motivated Online Benchmark for Image Matting[CVPR-2019]
