CV Chapter 5 Local Features and Matching
Questions about the lecture 'Computer Vision' of the RWTH Aachen Chapter 5 Local Features and Matching
Questions about the lecture 'Computer Vision' of the RWTH Aachen Chapter 5 Local Features and Matching
Set of flashcards Details
| Flashcards | 55 |
|---|---|
| Language | English |
| Category | Computer Science |
| Level | University |
| Created / Updated | 04.02.2017 / 20.02.2017 |
| Weblink |
https://card2brain.ch/box/20170204_cv_chapter_5_local_features_and_matching
|
| Embed |
<iframe src="https://card2brain.ch/box/20170204_cv_chapter_5_local_features_and_matching/embed" width="780" height="150" scrolling="no" frameborder="0"></iframe>
|
What are the motivations for local feature representations? [2]
1. Global representations have limitations
2. Local representations have increased robustness to occlusion, articulation and intra-category variation
Name two applications of local feature representations? [2]
1. Analyze if two images display same content from different point of view
2. Image stitching
What are the steps of image stitching? [3]
1. Detect feature points
2. Find corresponding pairs matching local descriptors
3. Align images
What are the steps of feature point detection (step one) of image stitching? [3]
1. Define region around keypoint
2. Extract and normalize region content
3. Compute local descriptor
What are the steps of pair matching (step two) of image stitching?
Find same points // Correctly find corresponding one
What are local feature detectors in recent applications in computer vision?
Basic building block
What are the requirements for local feature detectors? [5]
1. Regions are repeatable and accurate
2. Invariant to translation, rotation and scale changes
3. Features have locality making them robust to occlusion and clutter
4. Provide sufficient number of distinct regions
5. Operations must be close to real-time performance
Name available local feature detectors? [7]
1. Hessian and Harris
2. Laplacian, DoG
3. Hessian-/Harris-Laplace
4. Hessian-/Harris-Affine
5. EBR and IBR
6. MSER
7. Salient Regions
What are the goals of keypoint localization for local feature detectors? [3]
1. Repeatable detection
2. Precise localization
3. Interesting content
What are the basic techniques for keypoint localization for local feature detectors? [3]
1. Look for 2D signal changes
2. Around corner two or more dominant gradients exist
3. Slide small window
What are the three cases in keypoint localization for local feature detectors? [3]
1. Flat with no changes in all directions
2. Edge with no changes along edge direction
3. Corner with significant change in all directions
What are the definitions of the intensity shift change E(u,v) for the harris detector? [7]
1. E(u,v) = Sum_x,y w(x,y)[I(x+u,y+v)-I(x,y)]²
2. E(u,v) = [u,v]*M*[u;v]
With
3. w(x,y) is window function
4. I(a,b) is intensity function
5. M = Sum_x,y w(x,y)*[Ix²,IxIy;IxIy,Iy²]
6. M = [SumIx²,SumIxIy;SumIxIy,SumIy²]
7. M = Sum [Ix;Iy]*[Ix,Iy]
What holds for the window function for the harris detector? [3]
1. Not rotation invariant if rectangular
2. Rotation invariant with Gaussian g(sigma)
3. Not scale invariant
What is the definition for an axis-aligned corner for the harris detector? [2]
1. M = [l1,0;0,l2]
2. If one of l1 or l2 is close to zero → no corner
What is the definition for an general corner for the harris detector? [4]
1. M = R^-1*[l1,0;0,l2]*R // Eigenvalue decomposition
2. R = det(M)-a*trace(M)² = l1*l2 – a(l1+l2)²
3. Faster with no eigenvalues and a\in[0.04,0.06]
4. Visualization of M as ellipse
What are the characteristics of the visualization of an general corner for the harris detector? [3]
1. Corner if l1 l2 large and equal
2. Edge if l1>>l2 or l2>>l1
3. Flat if l1 l2 small and equal
What is the definition of the second moment (so-called autocorrelation) matrix for the harris detector? [2]
1. M(sI,sD)=g(sI)*[Ix²(sD),IxIy(sD);IxIy(sD),Iy²(sD)]
2. R = det[M(s1,s2)]-a[trace(M(s1,s2))]² = g(Ix²)g(Iy²)–[g(IxIy)]²-a[g(Ix²)+g(Iy)²]²
What are the characteristics of the hessian detector? [2]
1. Hessian(I)=[Ixx,Ixy;Ixy,Iyy] // 2nd derivativs
2. Hessian determinant det(Hessian(I)) = IxxIyy – Ixy²
What is the naive approach of region localization?
Naive approach vary patch size on one side is inefficient
What should region localization functions be?
Scale invariant
What is an example for region localization functions? [2]
1. Local maximum of average intensity for automatic scale selection
2. LoG (blob detector) as signature function
What is the definition of the characteristic scale for region localization functions?
Scale which produces peak of Laplacian response
What are the characteristics of Harris-/Hessian-Laplace? [2]
1. Multiscale Harris/Hessian detector
2. Set scale based on Laplacian
What is the simple local descriptor?
Region as vectors
What are the characteristic of simple local descriptor? [2]
1. Intensities within a patch
2. Shift invariant
What are the characteristic of the scale invariant feature transform (SIFT)? [3]
1. One patch divided into 4x4 sub-patches
2. Histogram of gradient orientations with 8 reference angles
3. Real time applicable
What is the scale invariant feature transform (SIFT) robust to? [2]
1. Robust to up to ~60° viewpoint rotation
2. Robust to changes in illumination
What are the parameters of scale invariant feature transform (SIFT)? [4]
1. n 2D points for positions
2. n scale parameters for size
3. n orientation parameters for angle
4. n 128-dimensional descriptors for gradient orientation
What are the characteristic SURF? [2]
1. Fast approximation of SIFT idea 6 times faster
2. Using 2D box filters and integral images
What are the applications of local descriptors? [6]
1. Wide baseline stereo
2. Motion tracking
3. Panoramas or mosaicing
4. Mobile robot navigation
5. 3D reconstruction
6. Recognition
What are the advantages of local descriptors? [4]
1. Critical to find local regions for multi-view matching
2. Complexity reduction via selection of distinctive points
3. Image description without segmentation
4. Robust to clutter, occlusion, view changes, noise, blur, etc
What are the two possibilities of recognition with local features? [2]
1. Warping and 2. alignment
What is the idea of warping for recognition with local features? [2]
1. Given image I and transformation T compute new image I’
2. T(p) = p’ for every pixel p with M*p = p’
What are the characteristic of 2D affine transformations (AT)? [2]
1. M is 3x3 and last row [0,…,0,1]
2. Parallel lines remain parallel
What are the characteristic of projective transformations (PT)? [4]
1. M is 3x3 with full
2. Combination of AT and projective warps
3. Parallel lines do not necessarily remain parallel
4. Called homography
What are the characteristic of homogeneous coordinates regarding homography? [2]
1. H*p = p’ with p=(x;y;1) and p’=(wx’;wy’;w) or p’=(x’;y’;z’)
2. Set scale to one H_33 reducing free parameters
What are the characteristic of image coordinates regarding homography?
1/z’ *p’=p’’ with p’=(x’;y’;z’) and p’’=(x’’;y’’;1)
What are the equations to solve the homography? [5]
1. x_A1 = H_11*x_B1+H_12*y_B1+H_13/H_31*x_B1+H_32*y_B1+1
2. H_11*x_B1+H_12*y_B1+H_13-x_A1*H_31*x_B1-x_A1*H_32*y_B1-X_A1=0
3. y_A1 = H_21*x_B1+H_22*y_B1+H_23/H_31*x_B1+H_32*y_B1+1
4. H_21*x_B1+H_22*y_B1+H_23-y_A1*H_31*x_B1-y_A1*H_32*y_B1-y_A1=0
5. Solve A*h=0 with A=U*D*V^T
What is the idea of alignment for recognition with local features? [4]
1. Given images I I’ computing transformation T
2. Find a and b for p*a + b = p’
3. With n points exact solution
4. With more than n points min|Ax-B|² // Least-squares minimization
What is the problem regarding outliers for recognition with local features? [4]
1. Erroneous matched pair
2. Feature point that is noise
3. Hurt quality of parameter estimates
4. In practical situations percentage of outliers>90%
