Learning Pixel-wise Alignment for Unsupervised Image Stitching
Qi Jia
Dalian University of Technology
jiaqi@dlut.edu.cn
Xiaomei Feng
Dalian University of Technology
xiaomeifeng19@gmail.com
Yu Liu
∗
Dalian University of Technology
liuyu8824@dlut.edu.cn
Xin Fan
Dalian University of Technology
xin.fan@dlut.edu.cn
Longin Jan Latecki
Temple University
latecki@temple.edu
ABSTRACT
Image stitching aims to align a pair of images in the same view.
Generating precise alignment with natural structures is challeng-
ing for image stitching, as there is no wider eld-of-view image
as a reference, especially in non-coplanar practical scenarios. In
this paper, we propose an unsupervised image stitching frame-
work, breaking through the coplanar constraints in homography
estimation, yielding accurate pixel-wise alignment under limited
overlapping regions. First, we generate a global transformation by
an iterative dense feature matching combined with an error control
strategy to alleviate the dierence introduced by large parallax. Sec-
ond, we propose a pixel-wise warping network embedded within a
large-scale feature extractor and a correlative feature enhancement
module to explicitly learn correspondences between the inputs,
and generate accurate pixel-level osets upon novel constraints
on both overlapping and non-overlapping regions. Notably, we
leverage the pixel-level osets in the overlapping area to guide the
adjustment in the non-overlapping area upon content and structure
consistency constraints, rendering a natural transition between two
regions and distortions suppression over the entire stitched image.
The proposed method achieves state-of-the-art performance that
surpasses both traditional and deep learning approaches by a large
margin. It also achieves the shortest execution time and has the
best generalization ability on the traditional dataset.
CCS CONCEPTS
• Computing methodologies → Computer vision;
KEYWORDS
image stitching, pixel-wise alignment, homography estimation
ACM Reference Format:
Qi Jia, Xiaomei Feng, Yu Liu, Xin Fan, and Longin Jan Latecki. 2023. Learning
Pixel-wise Alignment for Unsupervised Image Stitching. In Proceedings of
∗
Corresponding author: Yu Liu.
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for prot or commercial advantage and that copies bear this notice and the full citation
on the rst page. Copyrights for components of this work owned by others than the
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specic permission
and/or a fee. Request permissions from permissions@acm.org.
MM ’23, October 29-November 3, 2023, Ottawa, ON, Canada
© 2023 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 979-8-4007-0108-5/23/10.. . $15.00
https://doi.org/10.1145/3581783.3612298
(b) Stitched image of global alignment
(c) Stitched image of pixel-wise alignment
Network
Global alignment
(a) Global alignment vs. Pixel-wise alignment
Pixel-wise alignment
1 1 1
2 2 2
'
'
...
x x x
x x x
Offsets of four points
Pixel-wise offset
Figure 1: Global alignment vs. Pixel-wise alignment. (a) Il-
lustration of the dierence between global and pixel-wise
alignment in principle. Global alignment applies a direct
linear transformation (DLT) to approximate the oset for
all the pixels, while our pixel-wise alignment executes un-
uniform transformations for each pixel. (b) and (c) compare
two kinds of stitching results. Pixel-wise alignment achieves
superior results in image and artifact suppression compared
to global alignment, as shown in the zoomed-in regions.
the 31st ACM International Conference on Multimedia (MM ’23), October 29-
November 3, 2023, Ottawa, ON, Canada. ACM, New York, NY, USA, 9 pages.
https://doi.org/10.1145/3581783.3612298
1 INTRODUCTION
Image stitching aims to estimate an accurate transformation be-
tween a pair of images and align them in the same view. It has
been a well-studied topic with widespread applications [
38
] such
as panorama on smartphones [
42
], robot navigation [
7
], and virtual
reality [
1
,
18
]. However, generating high-quality stitched images
in various practical scenarios is still challenging, especially when
there is no wider eld-of-view image as a reference.
Homography transformation [
5
,
9
,
44
] is the most widely used
image stitching model, that leverages the feature correlation in
overlapping regions as constraints to estimate a global homogra-
phy matrix [
30
], and transform the whole target image to the view
of the reference image (see the global warping part in Fig. 1 (a)).
Most existing methods estimate the global homography by assum-
ing the whole scene is coplanar, leading to severe misalignment