Jan 24, 2017

Semantic Segmentation

  1. Semantic Segmentation with Second Order Pooling (CVPR'12)
    • Generate bottom-up multiple obj-vs-background region using binary graph-cuts
    • Generate second order pooling features on each region
      • extract dense SIFT (dense points inside a region)>create covariance matrix (hence second order)->take max/avg on the covariance matrix
    • Learn simple linear-svm classifier to classify region
    • Semantic Segmentation done by sequentially overlaying the classified regions
    • Does better than Berkeley vision classification on PASCAL VOC challenge. 
    • No codebook required to learn
    • Emphasizes pooling operations
    • Unlike first order max/average pooling, emphasizes second order pooling which results into better features (experimentally)
    • Faster learning/classification due to simple classifier unlike kernel-svms
  2. Perceptual Organization (CVPR'13)
  3. Fully Convolutional Network (CVPR'15)
  4.  Rest

No comments: