Shape-from-Template, or SfT in short, is a model-based approach to 3D reconstruction from the field of computer vision.
The general scope of 3D reconstruction is to infer quantitative 3D information from 2D images.
Even though it has analogies with stereopsis in the human visual system, 3D reconstruction in computer vision is only concerned with finding computational solutions.
There are many approaches to 3D reconstruction: SfT works with a single input image at a time and proceeds by fitting a deformable 3D object model to this image.
This model is represented by a 3D shape and a deformation law, which form the fundamental required elements for the object template.
SfT can also process recorded or live video streams.
The latter is a common case in SfT's augmented reality applications and requires realtime processing.
Processing video streams can be achieved trivially by applying single-image SfT to each video frame.
However, most videos exhibit temporal continuity which may be exploited by SfT to improve performance in several ways.
These include speeding up and robustifying the computation and increasing accuracy by smoothening the computed stream of 3D shape models.
The fitting fails if the object is not present in the image and SfT then answers that the object is absent.
The deformation law defines the constraints that any deformation of the object performed in 'regular' usage conditions must satisfy.
The template may contained any other prior knowledge on the object potentially useful for SfT, such as the object's reflectance function (known as the BRDF).
The scope of SfT is objects whose deformation space may be large but for which a physics-based deformation law is known.
SfT works by finding correspondences between the template model and the input image.
This is the registration step.
The desired 3D shape is then inferred from the registration.
This is the shape inference step.
An SfT algorithm may solve only one of these two steps or may solve both, sequentially or simultaneously.
