Face Swapping in Python Using dlib, OpenCV, and Procrustes Alignment

This tutorial explains how to build a concise (~200‑line) Python script that swaps faces between two images by automatically detecting facial landmarks, aligning the faces, correcting color mismatches, and blending the result.

1. Detect facial landmarks with dlib – The script loads the dlib shape predictor and uses detector = dlib.get_frontal_face_detector() and predictor = dlib.shape_predictor(PREDICTOR_PATH) . The get_landmarks(im) function returns a 68×2 matrix of (x, y) coordinates for each face, raising TooManyFaces or NoFaces exceptions when appropriate.

2. Align the faces using Procrustes analysis – The transformation_from_points(points1, points2) function implements orthogonal Procrustes: it recenters the point sets, scales them by their standard deviations, and computes the optimal rotation via singular value decomposition. The resulting affine matrix M is used with OpenCV’s warpAffine to map the second image onto the first.

3. Color correction – Because lighting and skin tone differ, the correct_colours(im1, im2, landmarks1) function applies Gaussian blur (controlled by COLOUR_CORRECT_BLUR_FRAC = 0.6 ) to both images, then scales the second image’s colors to match the first on a per‑pixel basis.

4. Create face masks and blend – The script defines landmark groups (eyes, eyebrows, nose, mouth) and builds convex hull masks with draw_convex_hull and get_face_mask . Masks are feathered ( FEATHER_AMOUNT = 11 ) and combined using numpy.max . The final output is computed as: output_im = im1 * (1 - combined_mask) + warped_corrected_im2 * combined_mask

5. Full implementation – The article provides the complete source code, including imports ( cv2 , dlib , numpy ), constant definitions, helper classes, and the main execution block that reads two images, computes the transformation, applies masking, color correction, and writes the result to output.jpg .

By following the step‑by‑step guide and using the provided code, readers can understand and reproduce a practical face‑swap pipeline that combines computer‑vision techniques with linear‑algebraic alignment.