Description:
Modeling a free-form
3D-surface from a single view has been a widely pursued problem. The
existing schemes are either fully-automatic shape-from-X techniques or
involve adept interaction from the user but little or no geometric
(photometric) basis. We propose a novel scheme of interactive modeling
of free-form lambertian surfaces where the solution obtained is
consistent with the Shape from Shading model. To this end, a
reinforcement learning based scheme has been adopted which allows user
intervention at any stage of the algorithm to guide the SFS solution to
a global minimum. The frame-work allows the incorporation of various shading and
illumination models as well as different camera models.
Results:
|
|
Input Image
of a Sphere
|
Reconstructed
Sphere
|
(Click
here for an animation)
|
|
Input Image
of a Cone
|
Reconstructed
Cone
|
(Click
here for an animation)
|
|
Input Image
|
CALA
placement
|
|
|
Surface
before interaction
|
Surface
after interaction
|
|
|
Novel View
of Original Vase
|
Reconstructed
Vase
|
|
|
Mozart Image
|
Hierarchical
Levels
|
|
|
Novel view
of original face
|
Reconstructed
face
|
|
|
Input Image
of a real Vase
|
Reconstructed
Vase
|
(Here as the input image shows significant specularity,
the hybrid shading model (Sepcular + Diffuse) was used in the reconstruction process. Click
here for an animation of the solution evolution process. Alternatively, click here to see an animation of the reconstructed Vase. )
|
|
Input Image
of a Capsicum
|
Reconstructed
Capsicum
|
(Here again,
the hybrid shading model (Sepcular + Diffuse) was used in the reconstruction process.)
|
|
Input Face Image
|
Reconstructed
face (click for video)
|
(Here again,
the hybrid shading model (Sepcular + Diffuse) was used in the reconstruction process.)
Report:
On Learning
Shapes from Shades
Subhajit
Sanyal, Mayank Bansal, Prem Kalra, Subhashis Banerjee
Technical Report 2004
|
