1. Abstract
   1. goal
      1. recover a coherent bump map from a single texture image
2. Introduction
   1. problems
      1. most textures are not limited to color variations
         1. bump mapping[1]
         2. 3D and geometric textures[7]
         3. bi directional texture functions[3,4,10]
         4. polynomial textures[11]
         5. relief texture mapping[12]
   2. shape from shading(SFS)
      1. Horn and Brooks [9]
      2. Zhang et al.
   3. approach of this paper
      1. use segementation to divide the texture into basic visual components
      2. identify different color zones
         1. shadows
         2. specular highlights
      3. identify "texture elements"
         1. spots
         2. bricks
         3. grains
         4. etc
      4. analyze the classified structures
   4. linearly separate the relief into two distinct scales
      1. large scale
         1. addressed by using the identified texture structures
      2. small scale
         1. addressed using a filter computed accoring to the light source direction
3. Texture and lighting conditions
   1. certain number of assumptions
      1. image and the lighting condition
         1. lighting is uniform over the entire image
         2. one white light source is assumed dominant and sufficiently far away to retrieve a constant lighting direction
         3. the observer is assumed perpendicular to the image plane
      2. the nature of the texture
         1. the relief can be entirely characterized by a height map
         2. there are no visible environment reflections on the texture
         3. the texture is composed of individual arbitrarily shaped bumps
   2. consider a number of important effects
      1. shadows
      2. specular highlights
      3. different hues
      4. roughness characteristics
         1. two roughness scales
            1. small scale relief
            2. correspond to noise like irregularities
            3. shadowing effefcts realted
            4. large scale relief
            5. has a relatively large amplitude
            6. also produce self shadows & specular highlights
            7. can be characterized by a single elevation curve (1D)
      5. etc.
4. Small Scale relief recovery
   1. design a simple filter based technique
      1. two assumptions
         1. noisy textures are visually alike
         2. the shading due to a unique directional light source modifies in two ways the frequency domain of the original bump map
      2. filter
         1. low pass filter(e.g. Gaussian)
5. Large Scaled relief recovery
   1. Identifying large scale structures
      1. based on training sets
      2. similar to Premoze et al.[15]
      3. obtained two zones after segmenting and the training set was composed of only two selected pixels
         1. example
      4. identifying the "bumps"
         1. obtain two types of segmentation
            1. color strctures
            2. bump shapes
            3. let users set a Boolean value for each
         2. assume that the relief of bump structures is related to theire shapes and can be characterized by a single elevation curve
         3. compute a large scale bump map simply using the distance transform
            1. based on erosion
   2. computation of an "unshaded" color map
      1. blur the segmented images
6. Combing large scale and small scale relief
   1. example
7. Conclusions and future directions
   1. constrains
      1. many assumptions
         1. the large scale relief could be characterized by a single curve
      2. cannot be used for reflectance recovery