1. abstract
   1. appearance-based method to model tree bark surfaces from a single image
   2. feature specification
      1. use texton channel analysis(the mesostructures)
      2. a variant of common bark features
         1. ironbark
         2. vertical and horizontal fractures
         3. tessellation
         4. furrowed cork
         5. lenticels
   3. input
      1. a bark image
   4. output
      1. a textured height field
2. Introduction
   1. surface geometry
      1. high level: meshes represent gross shape
      2. medium level: the mesostructure
      3. low level: bi-directional reflectance functions (BRDFs)
   2. two main objectives
      1. can capture a variety of bark features
      2. easy to use
   3. key ingredient
      1. the texton channel analysis
         1. proposed by Malik et al. [16]
         2. the bark image is segmented into texton channels
   4. interactive texture editing technique
      1. Brooks and Dodgson [2]
3. Related Work
   1. uses textures
      1. Oppenheimer [18]
      2. Hart and Baker [8]
   2. physics-based
      1. Federl and Prusinkiewicz [6]
      2. Hirota et al.[9]
      3. Lefebvre and Neyret[13]
      4. Terzopoulos and Fleisher[22]
   3. image-based
      1. Liu et al [15]
      2. Rushmeier et al [21]
      3. Dischler et al. [5]
      4. Leclerc et al.[12]
      5. require multiple images with known illumination.
4. Bark Modeling
   1. framework
      1. framework of our approach
   2. bark features
      1. ironbark
      2. fracture
      3. tessellation
      4. furrowed cork
      5. lenticel
      6. examples
         1. Bark features
   3. Texton Channels
      1. the texton channel analysis
         1. proposed by Malik et al. [16]
         2. filter
            1. two phases
            2. even
            3. odd
            4. three scales
            5. six orientations
         3. cluster
            1. For each pixel of I_0 , the filter response is a 36-dimensional data vector.
            2. clustered using the K-means algorithm
            3. The resulting K-means centers are the textons of I_0
            4. we set the number of K-means centers to K = 25
            5. example
            6. All 25 Texton channels(K = 25)
            7. texton channels
      2. interactly merge channels
         1. merging texton channels
   4. Height Field Construction
      1. UI for building the height field H0
         1. UI
      2. Starting with an H0=0 everywhere
      3. interactively edits the height values of H0
         1. propagation
         2. weight mask
            1. calculated by distance
            2. other ways
            3. same weight
            4. according to the grey-scale value of I0
   5. Texture Map Construction
      1. compute the texture map T0 from H0
         1. construct the texture T0 by dividing the image I0 by N · L on a per-pixel basis
5. Experimental Results
6. Discussion and Future Work
   1. potential improvement
      1. start with a better initial guess of H0
   2. weakness
      1. difficult in modeling "stringy bark"