AI summary by TwinMind
Lecture 10: Characteristic Strip Expansion, Shape from Shading, Iterative Solutions
MIT OpenCourseWare
1h 26m 33s
Course Announcements
- Three homework problems completed
- First take-home quiz next week
- Counts twice as much as homework
- No collaboration allowed (unlike homework)
- 686 students: proposal due on 22nd
- Short description of term project
- Machine vision problem implementation (Windows, Matlab, Android, etc.)
- Can be theoretical/mathematical solution
- Can use OpenCV but must add original contribution
- Email for project ideas/sources
Term Project Examples
- Time to contact implementation on Android phone
- Android Studio with dummy project shell provided
- Subpixel edge detection methods implementation
- Use Matlab or convenient platform
- Shape from shading with different reflectance maps
- More theoretical/mathematical approach
- Examples to be posted on Stella website
Course Structure Overview
- First part: image projection, perspective projection, motion analysis
- Current focus: image brightness measurements and shape from shading
- Next: industrial machine vision methods using patents
- Patents used instead of papers (industry doesn't publish for credit)
- Learn patent language for entrepreneurs/startups
- Transition from partial differential equations to practical applications
Image Irradiance Equation Foundation
- Brightness at image point = reflectance map of surface orientation
- Depends on: illumination, surface material, geometry (surface orientation)
- Local relationship: image brightness depends on corresponding object point
- Reflectance map summarizes detailed reflecting properties (BRDF)
- BRDF: 4 parameters → Reflectance map: 2 parameters
- Orthographic projection assumption (simplifies mathematics)
- Corresponds to telecentric lens or very far distance
- Like looking at moon with small visual angle
Hapka Model Solution Review
- Moon/rocky planet reflecting properties
- Can integrate in ecliptic plane direction only
- Characteristic curves with update rules:
- x, y coordinate changes
- Height change: dz = p·ps + q·qs
- Direct relationship: E²·rs - 1 = p·ps + q·qs
- March along curves adjusting x, y, z at each point
- Need initial curve for boundary conditions
- Can go bidirectionally from initial curve
General Shape from Shading Method
- Five coupled ordinary differential equations
- Variables: x, y, z, p, q (position + orientation)
- Characteristic strips: curves + surface orientation
- Update rules using chain rule for p, q changes
- Need second derivatives (Hessian matrix): r, s, t
- More unknowns than equations initially
Breakthrough: Using Image Irradiance Equation
- Brightness gradient: ∇E = [rp, rq] · H · [δx, δy]
- Pattern matching approach: δx = ε·rp, δy = ε·rq
- Eliminates need to solve for Hessian matrix H
- Direction determined by reflectance map gradient
- Five differential equations:
- dx/dψ = ε·rp
- dy/dψ = ε·rq
- dp/dψ = ε·ex
- dq/dψ = ε·ey
- dz/dψ = ε·(p·rp + q·rq)
Step Size Control Methods
- Constant step size in height (z-direction)
- Divide by (p·rp + q·rq)
- Move from contour to contour
- Constant step size in image plane
- Divide by √(rp² + rq²)
- Fixed pixel increments
- Constant step size in 3D space
- Divide by √(rp² + rq² + (p·rp + q·rq)²)
- Constant brightness increments
- Divide by (ex·rp + ey·rq)
- Move between isophotes
Specific Reflectance Map Applications
- Hapka model: reduces to original special case
- SEM model: R = f(√(p² + q²))
- Update direction = gradient direction
- Simple uphill/downhill movement
- Lambertian: more complex due to square root terms
Parallelization and Implementation
- Base characteristics: projection into image plane
- Solutions along curves are independent
- Parallelizable computation (not fully parallel)
- Wave front propagation approach
- Interpolate/merge characteristics as needed
- Monitor spacing between curves
Initial Conditions Problem
- Need initial curve for solution
- Worse: need initial strip (position + orientation)
- Two constraints available:
- Image irradiance equation: E(x,y) = R(p,q)
- Curve constraint: dz/dθ = p·dx/dθ + q·dy/dθ
- Two equations, two unknowns (p, q) - solvable
- Still need initial curve (problematic)
Stationary Points Solution
- Occluding boundary: known orientation but infinite slope
- Global extrema in reflectance map
- Lambertian: maximum brightness point
- SEM: minimum brightness point
- Unique surface orientation determination
- Problem: stationary points don't move (rp = rq = 0)
Local Shape Estimation
- Move slightly away from stationary point
- Use second derivatives of brightness
- Construct curved surface patch around stationary point
- Example: z = x² + 2y² surface
- Gradient of image gives curvature information
- Start solutions from constructed patch boundary
