Python np.uint8（）使法向量计算错误

所以，我想通过对深度图像应用梯度来计算边缘的法向量。当我出于某种原因（第一次尝试）使用np.uint8（）更改矩阵图像时，法线指向错误的方向，如下图所示。在法线函数中，我使用ndimage.sobel来计算梯度。

当我从生成的实矩阵（第二次尝试）更改输入图像时，法向量计算正确，如图所示。有人能解释为什么会这样吗？

---

原始数据具有

dtype（'float64'）

，其中

max=1.13

和

min=-0.11

。因此，转换为

dtype（'uint8'）

会大大降低精度。有关详细信息，请参阅。谢谢您的回答。

#1st try
sm_im = np.zeros((100,100))
sm_im[30:-30, 20:-20] = 1
sm_im = nd.rotate(sm_im, 25, mode='constant')
sm_im_copy = np.uint8(sm_im) #HERE IS THE np.uint8
sm_edge_im = cv2.Canny(sm_im_copy, 1, 5)
normals_pm, edge_pm = normals(sm_im_copy, sm_edge_im) #PAY ATTENTION HERE
cv2.imwrite(r'C:\Users\Adam Syammas Zaki P\Documents\Assignment\Grasp Planning\Simulation\Image\simple_debug_edge.png', sm_edge_im)
save_arr = r'C:\Users\Adam Syammas Zaki P\Documents\Assignment\Grasp Planning\Simulation\Image\sm_normal_debug.png'
depth = Image.open(r'C:\Users\Adam Syammas Zaki P\Documents\Assignment\Grasp Planning\Simulation\Image\simple.png')
draw_normals(edge_pm, normals_pm, depth, save_arr)

#2nd try
sm_im = np.zeros((100,100))
sm_im[30:-30, 20:-20] = 1
sm_im = nd.rotate(sm_im, 25, mode='constant') #USING THE ORIGINAL MATRIX
sm_im_copy = np.uint8(sm_im)
sm_edge_im = cv2.Canny(sm_im_copy, 1, 5)
normals_pm, edge_pm = normals(sm_im, sm_edge_im) #PAY ATTENTION HERE
cv2.imwrite(r'C:\Users\Adam Syammas Zaki P\Documents\Assignment\Grasp Planning\Simulation\Image\simple_debug_edge.png', sm_edge_im)
save_arr = r'C:\Users\Adam Syammas Zaki P\Documents\Assignment\Grasp Planning\Simulation\Image\sm_normal_debug.png'
depth = Image.open(r'C:\Users\Adam Syammas Zaki P\Documents\Assignment\Grasp Planning\Simulation\Image\simple.png')
draw_normals(edge_pm, normals_pm, depth, save_arr)