Region Proposal Algorithms

• Use segmentation to identify prospective objects
• Important to have a very high recall → Missing the actual regions severely impacts the detection rate
• Selective Search is the most commonly used algorithms as it is fast and has high recall

Selective Search

• A region proposal algorithm based on computing hierarchical grouping of similar regions based on color, texture, size and shape compatibility
• How selective search work with oversegmented image as initial input

• How do we calculate similarity between two regions?

  • Color similarity

    • color histogram of 25 bins is calculated for each channel of the image and channels are all concatenated resulting in a 25*3=75 dimensional color descriptor

    • Color similarity of two regions is based on histogram intersection

      

  • Texture Similarity

    • Extract Gaussian derivatives at 8 orientations for each channel
    • For each orientation and for each color channel, 10-bin histogram is computed resulting in a 1083=240 dimensional texture descriptor
    • Texture similarity of two regions is based on histogram intersections

    

  • Size Similarity

    • Encourage smaller regions to merge early
    • Ensure region proposals at all scales are formed at all parts of image

    

  • Shape Compatibility

    • Measures how well two regions fit into each other
    • If the two regions does not touch, they should not be merged

    

  • Final Similarity