[Image Interpretation] Object detection

Part-based Object Detection Method

Circle Hough Transform

• Radius-known case: the model parameter space is two dimensional: the coordinate of center.
• Procedure:
  
  • discretize model parameter space and create accumulation matrix;
  • calculate the pixel gradient (including magnitude and direction) of image, and choose those with high enough response as key point;
  • Each key point will vote for a set of model parameters. While models with strong image support will accumulate most votes.
• Radius-Unknown Case:
  
  • Either we have to add one dimension of the model parameter space – radius dimension and restart the discretization of model space and vote.
    
    • So curse of dimensionality: complexity increases exponentially with the number of model parameter
  • Instead, with the knowledge of the gradient direction, the center must locate on this line of direction – Thus transform the conical surface into a line in parameter space
• Generalized Hough Transform: with the introduction of predefined reference point and pre-calculated R-table.
  
  • The purpose of GHT is still construct model parameter space, which corresponding to the find the most supported reference point.
• Pros:
  
  • Points are processed indpendently: robustness to noise and occlusion
• Cons:
  
  • Non-object can also contribute to spurious peak in model space.

Implicit Shape Model

• Divided into two procedure:
  
  • Learning a object model:
    
    • Extract interest point, compute descriptor and offset to a reference point;
    • Cluster descriptors and offsets to form a visual dictionary
  • Object detection:
    
    • Extract interest point, compute descriptor
    • Find the most similar descriptor and vote for the possible reference point
• Non-maxima suppression: find the strongest response

Window-based Object Detection Method

• Sliding window: direct template matching - normalized cross-correlation
• Viola-Jones detector: for person face detection:
  
  • A very unbalanced tree
• Deformable part model: relative placement of part is allowed to change

• Window proposal:

  • Instead of searching all over the window, only search candidate wondiow which may contain object.
  • Also accompany image segmentation
  • Not suitable for diffuse or low-contrast object
  • few false positive but more false negative.

  Histogram of oriented gradient

  In the HOG feature descriptor, the distribution ( histograms ) of directions of gradients ( oriented gradients ) are used as features.

  • Procedure:
  • Calculate the gradient (including magnitude and direction) of each pixel. For 3 Channels image, we simply use the maximum magnitude among three channels and its corresponding direction (angle).
  • Divide the whole image patch into $8 \times 8$ subpatch. Note that the image patch has already been scaled a patch of $64 \times 128 \times 3$.
  • Each $8 \times 8$ subpatch creates a histogram. The horizontal axis of histogram is divided into 9 bins, which corresponding to 9 separated group of the direction of gradient(angle). Here the angle (direction of gradient) ranges from 0 to 180 degree.
  • Each pixel will contributes to its corresponding bin according to its direction. The magnitude of contribution is proportional to its magnitude.
  • Normalized histogram by forming a $16 \times 16$ block, to eliminate the effect of brightness variation. Then a vector of $36 \times 1$ is created.

Refernece

Histogram of Oriented Gradient