Please enable JavaScript.

Coggle requires JavaScript to display documents.

Computer Vision - Coggle Diagram

- - - - Hausdorff Distance
        Give edge pixel location A and B
        We calculate only the maximum distance
        by comparing all the point pairs.
      - Divide and Conquer
        Cell (3D matrix) is divided and pruned when possible and if not possible, further divide into smaller cells.
  - - - Keypoints searching
        
        Finding Keypoints (Orientation)
        Histogram of local gradient is used. Choose the direction with the most votes
        
        Finding Keypoints (Location, Scale)
        
        Maxima/Minima of Difference of Gaussian
        
        Gaussian blurring is applied for different scales, and their different is calculated afterward. For each differences, search for the minimum/maximum in all 3 direction, its neighboring pixels, the upper scale neighbor pixels, and lower scale neighbor pixels. These are possible candidate of key-points.
      - Signature/Descriptor
        
        Image gradient is sampled over 16x16 array
        
        Divide the array into 4x4 array with each bin can have 8 orientation (+45degree each)
- - - - Sobel Filter
      - Canny Edge Detector
        
        What we covered in program assignment
      - Zero Crossing Detector
        
        Uses 2nd derivative and pick the point where the change in value is the greatest (or a.k.a inflection point)
        
        Always forms contour
  - - - Say we have a line \(y = m_0 x + b_0\) in image space, this can be represented as point \((m_0, b_0)\) in hough space
        If, instead, we have a point \((x_0, y_0)\) in image space, this will correspond to the line in the hough space of all values satisfying \(y_0 = mx_0 + b\) (or using sin cos line)
- - - - Histogram Distance
        
        Minkowski-type Distance
        
        Quadratic-form Distance
        
        Earth Move's Distance
      - Layout Template
        Color histogram is not aware of pixel locations. Spatial information may be added using template images.
- - - - Connected Components: Once the threshold is applied, we can find the connected components
        
        Recursive Growing
        
        Parallel Growing
        
        Raster Scan with Efficient Merging
  - - - Minimal Cut
      - Normalized Cut
      - GrabCut (Craph Cut)