Tal Amir Advanced Topics in Computer Vision May 29 th, 2015 COUPLED MOTION- LIGHTING ANALYSIS
1.Surface reconstruction: Overview Structure from Motion Photometric Reconstruction Coupled Motion-Lighting Analysis 2.Light scattering models Lambertian reflectance BRDF 3.Shape from object motion 4.Generalizing to unknown lighting and BRDF 5.Shape from camera motion 6.Conclusion OUTLINE
SURFACE RECONSTRUCTION: OVERVIEW
Structure from motion: Images are taken from different locations Scene remains static Seek dense correspondence SURFACE RECONSTRUCTION: OVERVIEW
Structure from motion: Images are taken from different locations Scene remains static Seeking dense correspondence Brightness constancy assumption SURFACE RECONSTRUCTION: OVERVIEW
Structure from motion: Images are taken from different locations Scene remains static Seek dense correspondence Use it to recover depths Brightness constancy assumption SURFACE RECONSTRUCTION: OVERVIEW
Structure from motion: Images are taken from different locations Scene remains static Seek dense correspondence Use it to recover depths Brightness constancy assumption SURFACE RECONSTRUCTION: OVERVIEW Very simplistic assumption Requires Lambertian reflectance
Photometric reconstruction (stereo): Images of the scene are taken with different illuminations Scene and camera remain static Uses a light scattering model SURFACE RECONSTRUCTION: OVERVIEW
Basri, R., Jacobs, D., & Kemelmacher, I. (2007). Photometric stereo with general, unknown lighting. International Journal of Computer Vision, 72(3),
Photometric reconstruction (stereo): Few images of the scene are taken with different illuminations Scene and camera remain static Uses a light scattering model What about reconstruction with… Object motion? Camera motion? SURFACE RECONSTRUCTION: OVERVIEW
Coupled motion-lighting analysis Uses a light scattering model Takes object / camera motion into account Assumes small (differential) motion, which is known SURFACE RECONSTRUCTION: OVERVIEW
Why assume differential motion? If the motion is large… SURFACE RECONSTRUCTION: OVERVIEW
Why assume differential motion? …we need point correspondences. SURFACE RECONSTRUCTION: OVERVIEW
Why assume differential motion? If we have a correspondence for all points, we’ve already solved the problem. Therefore, we assume the camera rotation and translation are small. SURFACE RECONSTRUCTION: OVERVIEW
1.Surface reconstruction: Overview Structure from Motion Photometric Reconstruction Coupled Motion-Lighting Analysis 2.Light scattering models Lambertian reflectance BRDF 3.Shape from object motion 4.Generalizing to unknown lighting and BRDF 5.Shape from camera motion 6.Conclusion OUTLINE
1.Surface reconstruction: Overview Structure from Motion Photometric Reconstruction Coupled Motion-Lighting Analysis 2.Light scattering models Lambertian reflectance BRDF 3.Shape from object motion 4.Generalizing to unknown lighting and BRDF 5.Shape from camera motion 6.Conclusion OUTLINE
What happens when a beam of light hits a surface? LIGHT SCATTERING MODELS
What happens when a beam of light hits a surface? LIGHT SCATTERING MODELS Specular reflection
What happens when a beam of light hits a surface? LIGHT SCATTERING MODELS Specular reflection Diffuse reflections
What happens when a beam of light hits a surface? LIGHT SCATTERING MODELS Specular reflection Diffuse reflections Subsurface scattering There are other, more complex phenomena, such as interreflections and shadow casting.
LIGHT SCATTERING MODELS
Bidirectional Reflectance Distribution Function (BRDF): LIGHT SCATTERING MODELS
Bidirectional Reflectance Distribution Function (BRDF): Generalizes specular and Lambertian reflections. A narrower model: Symmetric BRDF. Rotation invariant. Depends only on the angles between the three vectors. LIGHT SCATTERING MODELS
Bidirectional Reflectance Distribution Function (BRDF): LIGHT SCATTERING MODELS
Bidirectional Reflectance Distribution Function (BRDF): Generalizes specular and Lambertian reflections. A narrower model: Symmetric BRDF. Rotation invariant. Depends only on the angles between the three vectors. Does not account for subsurface scattering. Only good for opaque surfaces. LIGHT SCATTERING MODELS
Bidirectional Reflectance Distribution Function (BRDF): LIGHT SCATTERING MODELS
Bidirectional Reflectance Distribution Function (BRDF): LIGHT SCATTERING MODELS
1.Surface reconstruction: Overview Structure from Motion Photometric Reconstruction Coupled Motion-Lighting Analysis 2.Light scattering models Lambertian reflectance BRDF 3.Shape from object motion 4.Generalizing to unknown lighting and BRDF 5.Shape from camera motion 6.Conclusion OUTLINE
1.Surface reconstruction: Overview Structure from Motion Photometric Reconstruction Coupled Motion-Lighting Analysis 2.Light scattering models Lambertian reflectance BRDF 3.Shape from object motion 4.Generalizing to unknown lighting and BRDF 5.Shape from camera motion 6.Conclusion OUTLINE
Basri, R., & Frolova, D. (2008): A two-frame theory of motion, lighting and shape Assumptions: Orthographic projection Known (differential) object motion Lambertian reflectance Possibly non-uniform albedo Known directional lighting SHAPE FROM OBJECT MOTION Basri, R., & Frolova, D. (2008, June). A two-frame theory of motion, lighting and shape. In Computer Vision and Pattern Recognition, CVPR IEEE Conference on (pp. 1-7). IEEE.
SHAPE FROM OBJECT MOTION
We need boundary conditions. At the boundary of the silhouette, we know the direction of the surface normal. SHAPE FROM OBJECT MOTION
Results SHAPE FROM OBJECT MOTION
1.Surface reconstruction: Overview Structure from Motion Photometric Reconstruction Coupled Motion-Lighting Analysis 2.Light scattering models Lambertian reflectance BRDF 3.Shape from object motion 4.Generalizing to unknown lighting and BRDF 5.Shape from camera motion 6.Conclusion OUTLINE
1.Surface reconstruction: Overview Structure from Motion Photometric Reconstruction Coupled Motion-Lighting Analysis 2.Light scattering models Lambertian reflectance BRDF 3.Shape from object motion 4.Generalizing to unknown lighting and BRDF 5.Shape from camera motion 6.Conclusion OUTLINE
Chandraker, M., Reddy, D., Wang, Y., & Ramamoorthi, R. (2013): What object motion reveals about shape with unknown BRDF and lighting A general framework for surface reconstruction from camera motion. GENERALIZING TO UNKNOWN LIGHTING AND BRDF Chandraker, M., Reddy, D., Wang, Y., & Ramamoorthi, R. (2013, June). What object motion reveals about shape with unknown BRDF and lighting. In Computer Vision and Pattern Recognition (CVPR), 2013 IEEE Conference on (pp ). IEEE.
Tells us when we can or cannot reconstruct surface. Treats both orthographic and perspective projections. Generalizes to complex unknown illumination. Does not assume Lambertian reflectance. Limiting assumptions: Known differential motion Symmetric BRDF Object is distant from camera and light sources GENERALIZING TO UNKNOWN LIGHTING AND BRDF
Results GENERALIZING TO UNKNOWN LIGHTING AND BRDF
Lighting is co-located with the camera: Characteristic curves Initial values GENERALIZING TO UNKNOWN LIGHTING AND BRDF
Unknown lighting: Characteristic curves Initial values GENERALIZING TO UNKNOWN LIGHTING AND BRDF
Results GENERALIZING TO UNKNOWN LIGHTING AND BRDF
Synthetic images: GENERALIZING TO UNKNOWN LIGHTING AND BRDF
Real images: GENERALIZING TO UNKNOWN LIGHTING AND BRDF
Real images: GENERALIZING TO UNKNOWN LIGHTING AND BRDF
1.Surface reconstruction: Overview Structure from Motion Photometric Reconstruction Coupled Motion-Lighting Analysis 2.Light scattering models Lambertian reflectance BRDF 3.Shape from object motion 4.Generalizing to unknown lighting and BRDF 5.Shape from camera motion 6.Conclusion OUTLINE
1.Surface reconstruction: Overview Structure from Motion Photometric Reconstruction Coupled Motion-Lighting Analysis 2.Light scattering models Lambertian reflectance BRDF 3.Shape from object motion 4.Generalizing to unknown lighting and BRDF 5.Shape from camera motion 6.Conclusion OUTLINE
Chandraker, M., (2014): What camera motion reveals about shape with unknown BRDF Handles the case of camera motion, while object and lighting remain static. This task is more complex than object motion. SHAPE FROM CAMERA MOTION Chandraker, M. (2014, June). What camera motion reveals about shape with unknown BRDF. In Computer Vision and Pattern Recognition (CVPR), 2014 IEEE Conference on (pp ). IEEE.
For orthographic projection: Depth cannot be recovered purely from camera motion. A quasi-linear PDE can be derived using 3 images, if we make some restrictive assumptions on the BRDF. SHAPE FROM CAMERA MOTION
For orthographic projection: Depth cannot be recovered purely from camera motion. A quasi-linear PDE can be derived using 3 images, if we make some restrictive assumptions on the BRDF. For perspective projection: Depth can be directly recovered from 4 images. An additional linear PDE can be recovered if we make some restrictive assumptions on the BRDF. SHAPE FROM CAMERA MOTION
Results GENERALIZING TO UNKNOWN LIGHTING AND BRDF
Orthographic projection (synthetic images): GENERALIZING TO UNKNOWN LIGHTING AND BRDF
Perspective projection (synthetic images): GENERALIZING TO UNKNOWN LIGHTING AND BRDF
Perspective projection (real images): GENERALIZING TO UNKNOWN LIGHTING AND BRDF
1.Surface reconstruction: Overview Structure from Motion Photometric Reconstruction Coupled Motion-Lighting Analysis 2.Light scattering models Lambertian reflectance BRDF 3.Shape from object motion 4.Generalizing to unknown lighting and BRDF 5.Shape from camera motion 6.Conclusion OUTLINE
1.Surface reconstruction: Overview Structure from Motion Photometric Reconstruction Coupled Motion-Lighting Analysis 2.Light scattering models Lambertian reflectance BRDF 3.Shape from object motion 4.Generalizing to unknown lighting and BRDF 5.Shape from camera motion 6.Conclusion OUTLINE
We have seen: Surface reconstruction from two frames Known directional lighting Lambertian reflectance Orthographic projection General framework for object motion Unknown illumination Symmetric BRDF Orthographic / perspective projection General framework for camera motion CONCLUSION