关于Python与Matlab中仿射变换（Affine Transform）结果不一致的技术咨询

Hey there! Let’s break down exactly why your MATLAB affine transform isn’t matching Python’s output, and walk through how to align them step by step.

Common Causes of Discrepancy #

Most differences boil down to subtle but critical differences in how each tool handles transform order, coordinate systems, and parameters:

1. Transform Matrix Order #

MATLAB applies matrix transformations in right-to-left order for homogeneous matrices. For your code, that means the sequence is:
Rotation → Scaling → Translation (since your combined matrix would be T = RTrans * RScale * RRot).

Python libraries like OpenCV follow the same matrix multiplication logic, but it’s easy to accidentally reverse the order when building transforms manually, leading to misaligned results.

2. Rotation Direction #

MATLAB’s standard rotation matrix uses counterclockwise rotation (per mathematical convention). If you use OpenCV’s cv2.getRotationMatrix2D() without adjusting, it defaults to clockwise rotation—this flips your result immediately.

3. Interpolation & Boundary Handling #

MATLAB’s imwarp defaults to bilinear interpolation and fills out-of-bounds pixels with 0. If Python uses a different interpolation method (e.g., nearest-neighbor) or boundary mode, your output will look different even if the transform math is correct.

4. Data Type & Normalization #

Your MATLAB code uses im2double() to convert the image to a 0-1 float range. If Python keeps the image in 0-255 uint8 format, interpolation results will not match.

Fix: Align Python Code to MATLAB’s Behavior #

Here’s a Python implementation that mirrors your MATLAB code exactly, ensuring identical results:

import numpy as np
import cv2

# 1. Load and normalize image to match MATLAB's im2double
image = cv2.imread('fig2.bmp', cv2.IMREAD_GRAYSCALE)
image = image.astype(np.float64) / 255.0  # Convert to 0-1 float

# 2. Reuse your exact transform parameters
Rot = 0.001981123573629  # Rotation in radians (counterclockwise)
XformScale = [1.022909263199710, 1.029291261036412]  # [x-scale, y-scale]
XformTrans = [-0.405901345569081, 0.456816768280493]  # [x-translate, y-translate]

# 3. Build identical homogeneous matrices as MATLAB
# Rotation matrix (counterclockwise, matches MATLAB)
c = np.cos(Rot)
s = np.sin(Rot)
RRot = np.array([
    [c, -s, 0],
    [s, c, 0],
    [0, 0, 1]
])

# Scaling matrix
RScale = np.eye(3)
RScale[0, 0] = XformScale[0]
RScale[1, 1] = XformScale[1]

# Translation matrix
RTrans = np.eye(3)
RTrans[0, 2] = XformTrans[0]
RTrans[1, 2] = XformTrans[1]

# 4. Combine matrices in the SAME order as MATLAB (right-to-left)
T = RTrans @ RScale @ RRot  # Equivalent to MATLAB's RTrans * RScale * RRot

# 5. Convert to OpenCV's 2x3 affine matrix format
affine_mat = T[:2, :]

# 6. Apply transform with matching interpolation/boundary settings
h, w = image.shape[:2]
transformed_image = cv2.warpAffine(
    image,
    affine_mat,
    (w, h),  # Keep output size identical to input
    flags=cv2.INTER_LINEAR,  # Bilinear interpolation (MATLAB default)
    borderMode=cv2.BORDER_CONSTANT,
    borderValue=0.0  # Fill edges with 0 (MATLAB default)
)

Key Checks to Verify #

• Double-check the matrix multiplication order: RTrans @ RScale @ RRot is critical to match MATLAB’s right-to-left application.
• Ensure you’re using counterclockwise rotation (our code directly mirrors MATLAB’s rotation matrix, so no angle adjustment needed).
• Confirm image normalization: converting to 0-1 float ensures interpolation matches MATLAB’s im2double() output.

内容的提问来源于stack exchange，提问作者Xufu Sun