These years we have been observing a tremendous resurgence of
interest in stereoscopic 3D. A variety of stereoscopic displays and
cameras are available. This brings in the demand for tools for authoring
and processing stereoscopic content. However, extending
existing tools to stereoscopic content is often non-trivial as stereoscopic
content has an extra dimension of disparity that needs to be
correctly taken care of to deliver a pleasant viewing experience.
This paper focuses on stereoscopic image warping. Warping is one
of the basic image processing techniques and a wide range of image
warping methods have been developed [Wolberg 1990]. Applying
the same warping to the left and right image of a stereoscopic image
can transform them consistently; however this straightforward
solution is often problematic. Figure 1 shows a simple 2D image
transformation: rotation. If we rotate the left and right image with
the same rotation matrix, we introduce vertical disparities, which
often bring “3D fatigue” to viewers [Mendiburu 2009]. Moreover,
the original horizontal disparity distribution is changed.
This paper presents a technique that extends existing image warping
algorithms to stereoscopic images. Our idea is to warp one
of the two images of a stereoscopic image using the user-specified
warping and warp the other to both follow the user-specified warping
and meet the disparity requirement. Our technique consists
of three steps. We first apply the user-specified warping to one
of the two images. Without loss of generality, we always warp
the left image using the user-specified warping. We then compute
the target disparity map according to the user-specified warping.
We consider that a good target disparity map should be consistent
with the warping applied to the input image and maintain the perceived
roundness of the image objects. For example, if an object is
stretched in the image space, it should also be stretched in depth. If
the disparity remains the same, the warped image may suffer from
the cardboarding artifacts where the perceived object becomes flattened
[Mendiburu 2009]. Based on this observation, we develop an
automatic disparity mapping technique that scales the local disparity
range according to how this region is warped. Thus, the target
disparity map is optimized to preserve