Method for synthetically relighting images of objects
View Patent ↗An image of an object from a known object class is synthesized by first obtaining reflectance fields for various training objects from the object class. A reflectance field model is defined for the object class using a combination of the reflectance fields of the training objects. The parameters of the reflectance field model are optimized to estimate a particular reflectance field of a particular object from the object class given one or more input images of the particular object. The particular reflectance field is fitted to the particular object, and then the new image of the particular object is synthesized by changing the illumination parameters of the particular fitted reflectance field model after the fitting.
1. A method for synthesizing a new image of an object from an object class, wherein the object class is known, comprising the steps of:
obtaining reflectance fields for training objects from the object class;
constructing a reflectance field model for the object class using a combination of the reflectance fields of the training objects, wherein the constructed reflectance field model is defined as a mean of the reflectance fields of the training objects plus a linear combination of principal components of the reflectance fields of the training objects;
fitting the reflectance field model to a particular object from the object class by optimizing parameters of the reflectance field model to estimate a particular reflectance field given one or more input images of the particular object; and
synthesizing the new image of the particular object by changing illumination inputs of the particular reflectance field.
2. The method of claim 1 , wherein the constructed reflectance field model is further defined as a product of a linear illumination field model and a separate linear texture model.
3. The method of claim 2 , wherein the illumination field model is defined as a linear combination of illumination fields of the training objects.
4. The method of claim 2 , wherein the illumination field model is defined as a mean of the illumination fields of the training objects plus a linear combination of principal components of the illumination fields of the training objects.
5. The method of claim 2 , wherein the texture model is defined as a linear combination of a set of texture images.
6. The method of claim 2 , wherein the texture model is defined as a mean of a set of texture images plus a linear combination of principal components of the set of texture images.
7. The method of claim 5 , wherein the set of texture images is derived from the reflectance fields of the training objects.
8. The method of claim 6 , wherein the set of texture images is derived from the reflectance fields of the training objects.
9. The method of claim 5 , wherein each texture image is obtained from a frontally lit image of an object from the object class.
10. The method of claim 6 , wherein each texture image is obtained from a frontally lit image of an object from the object class.
11. The method of claim 7 , wherein each texture image is obtained from the reflectance field of a training object by averaging the reflectance field over all illumination directions.
12. The method of claim 8 , wherein each texture image is obtained from the reflectance field of a training object by averaging the reflectance field over all illumination directions.
13. The method of claim 1 , wherein the reflectance fields are represented by tensor splines.
14. The method of claim 1 , wherein reflectance functions of the reflectance fields are represented by spherical harmonics.
15. The method of claim 1 , wherein reflectance functions of the reflectance fields are represented using a direct light combination method.
16. The method of claim 3 , wherein the illumination fields are represented by tensor splines.
17. The method of claim 4 , wherein the illumination fields are represented by tensor splines.
18. The method of claim 3 , wherein illumination functions of the illumination fields are represented by spherical harmonics.
19. The method of claim 4 , wherein illumination functions of the illumination fields are represented by spherical harmonics.
20. The method of claim 3 , wherein illumination functions of the illumination fields are represented by a direct light combination method.
21. The method of claim 4 , wherein illumination functions of the illumination fields are represented by a direct light combination method.
22. The method of claim 1 wherein the synthesizing further comprises:
applying quotient mapping to the new image.
23. The method of claim 1 , further comprising:
repeating the synthesizing with different illumination inputs to generate multiple new images; and
storing the new images along with the one or more input images to form an augmented gallery.
24. The method of claim 23 , wherein the augmented gallery is used for recognizing objects within the object class.
25. The method of claim 1 , further comprising:
repeating the fitting and synthesizing for more than one particular object from the object class.
26. The method of claim 1 , wherein the object class is faces.
27. The method of claim 1 , wherein at least one of the input images of the particular object is lit from a roughly frontal direction.
28. The method of claim 1 , wherein all of the input images of the particular object are pixelwise aligned with each other.