Key components: To design a very efficient computing algorithm for energy minimization. (1) Dynamic hierarchical structure (2) Simplified dynamics for graph change (3) Always steepest procedures for energy minimization.

   Key components: Learning-based similarity measure of different images by fusing many features.

  Probabilistic Boosting-Tree: Learning Discriminative Models for Classification, Recognition, and Clustering (paper)

     Key components: Learning discriminative models of large intra-class variation and inter-class similarity. (1) Hierarchical structure with decision tree + Boosting (2) Automatic clustering (3) Hidden variable augmentation (4) Overall posterior probability.

  Supervised Learning of Edges and Object Boundaries (paper)

     Key components: Push the learning side of the edge detection. (1) Automatically combine a large number of features across different scales and a variety of information.. (2) Implicitly take mid-level and context information into account. (3) Easy to train/test and fast to compute.

Detecting Object Boundaries Using Low-, Mid-, and High-Level Information (paper)

    Key components: Integrate low-, mid-, and high-level information. (1) Learning features from implicit ones to explicit ones. (2) Straightforward learning and efficient computing. (3) Useful for a wide class of objects.

  Feature Mining for Image Classification (paper)

     Key components: Automate feature design. (1) Principled way of taking account of computational complexity.  (2) Data-driven feature space and clustering. (3) Strategies for feature mining. (4) General system for object classification and detection.

  Generative Model Learning via Discriminative Approaches (paper)

     Key components: A unified view of generative and discriminative models. (1) Learning generative models progressively through discriminative approaches.  (2) Automatically handle a wide variety of patterns. (3) Improving efficiency and effectiveness of existing generative model learning. (4) Improving the modeling power of discriminative models.

  Image Segmentation by Data-driven Markov Chain Monte Carlo (paper)

  Parsing Images into Regions, Curves, and Curve Groups (paper)

  Image Parsing: Unifying Segmentation, Detection, and Object Recognition (paper)

     Key components: An integrated system for image segmentation. (1) Use discriminative models to guide generative model search.  (2) More robust than discriminative approaches. (3) More efficient than existing MCMC approaches. (4) Unified computing framework with different dynamics. (5) Interesting connections to biological vision systems. (6) Handles model selection and switch.

  Shape Matching and Registration by Data-driven EM (paper)

     Key components: A general system for shape matching.  (1) Use discriminative models in the EM to avoid local minimal  (2) More effective features in closed contour than shape context. (3) More efficient computing than soft assign.