Outline • S. C. Zhu, X. Liu, and Y. Wu, “Exploring Texture Ensembles by Efficient Markov Chain Monte Carlo”, IEEE Transactions On Pattern Analysis And Machine Intelligence, Vol. 22, No. 6, pp. 554-569, 2000 Limitations of Linear Representations • Linear representations do not depend on the spatial relationships among pixels – For example, if we shuffle the pixels and corresponding representations, then the classification results will remain the same • But in images spatial relationships are important May 29, 2016 Computer Vision 2 Image Features May 29, 2016 Computer Vision 3 Spectral Representation of Images • Spectral histogram – Given a bank of filters F(a), a = 1, …, K, a spectral histogram is defined as the marginal distribution of filter responses I(a ) (v) F (a ) * I(v) H (a ) I 1 (a ) ( z) δ ( z I (v)) |I| v H I ( H I(1) , H I( 2) ,, H I( K ) ) May 29, 2016 Computer Vision 4 Spectral Representation of Images - continued • An example of spectral histogram May 29, 2016 Computer Vision 5 Image Modeling - continued • Given observed feature statistics {H(a)obs}, we associate an energy with any image I as K (a ) Ε (I) |H I(a ) ( z ) H obs ( z ) |p a 1 z • Then the corresponding Gibbs distribution is q(I) 1 E (I) exp( ) ZT T – The q(I) can be sampled using a Gibbs sampler or other Markov chain Monte-Carlo algorithms May 29, 2016 Computer Vision 6 Image Modeling - continued Image Synthesis Algorithm • Compute {Hobs} from an observed texture image • Initialize Isyn as any image, and T as T0 • Repeat Randomly pick a pixel v in Isyn Calculate the conditional probability q(Isyn(v)| Isyn(-v)) Choose new Isyn(v) under q(Isyn(v)| Isyn(-v)) Reduce T gradually • Until E(I) < e May 29, 2016 Computer Vision 7 A Texture Synthesis Example Observed image May 29, 2016 Initial synthesized image Computer Vision 8 A Texture Synthesis Example Temperature Image patch Energy Conditional probability • May Energy and conditionalComputer probability of the marked 29, 2016 Vision pixel 9 A Texture Synthesis Example - continued • A white noise image was transformed to a perceptually similar texture by matching the spectral histogram May 29, 2016 Computer Vision 10 Average spectral histogram error A Texture Synthesis Example - continued • Synthesized images from different initial conditions May 29, 2016 Computer Vision 11 Texture Synthesis Examples - continued Observed image Synthesized image • A random texture image May 29, 2016 Computer Vision 12 Texture Synthesis Examples - continued Observed image Synthesized image • An image with periodic structures May 29, 2016 Computer Vision 13 Texture Synthesis Examples - continued Mud image Synthesized image • A mud image with some animal foot prints May 29, 2016 Computer Vision 14 Texture Synthesis Examples - continued Observed image Synthesized image • A random texture image with elements May 29, 2016 Computer Vision 15 Texture Synthesis Examples - continued Observed image Synthesized image • An image consisting of two regions – Note that wrap-around boundary conditions were used May 29, 2016 Computer Vision 16 Texture Synthesis Examples - continued Synthesized image Original cheetah skin patch • A cheetah skin image May 29, 2016 Computer Vision 17 Texture Synthesis Examples - continued Observed image Synthesized image • An image consisting of circles May 29, 2016 Computer Vision 18 Texture Synthesis Examples - continued Observed image Synthesized image • An image consisting of crosses May 29, 2016 Computer Vision 19 Texture Synthesis Examples - continued Observed image Synthesized image • A pattern with long-range structures May 29, 2016 Computer Vision 20 Object Synthesis Examples • As in texture synthesis, we start from a random image • In addition, similar object images are used as boundary conditions in that the corresponding pixel values are not updated during sampling process May 29, 2016 Computer Vision 21 Object Synthesis Examples - continued May 29, 2016 Computer Vision 22 Object Synthesis Examples - continued May 29, 2016 Computer Vision 23 Linear Transformations of Images • Linear transformations include – – – – – Principal component analysis Independent component analysis Fisher discriminant analysis Optimal component analysis They have been widely used to reduce dimension of images for appearance-based recognition applications • Each image is viewed as a long vector and projected into a set of bases that have certain properties May 29, 2016 Computer Vision 24 Principal Component Analysis • Defined with respect to a training set such that the average reconstruction error is minimized May 29, 2016 Computer Vision 25 Principal Component Analysis - continued May 29, 2016 Computer Vision 26 Eigen Values of 400 Eigen Vectors May 29, 2016 Computer Vision 27 Principal Component Analysis - continued Original Image May 29, 2016 Reconstructed using 50 PCs Computer Vision Reconstructed using 200 PCs 28 Principal Component Analysis - continued • Is PCA representation a good representation of images for recognition in that images that have similar principal representations are similar? – Image generation through sampling – Roughly speaking, we try to generate images that have the given coefficients along PCs May 29, 2016 Computer Vision 29 Principal Component Analysis - continued May 29, 2016 Computer Vision 30 Principal Component Analysis - continued May 29, 2016 Computer Vision 31 Difference Between Reconstruction and Sampling Reconstruction is not sufficient to show the adequacy of a representation and sampling from the set of images with same representation is more informational May 29, 2016 Computer Vision 32 Object Recognition Experiments • We compare linear methods in the methods including – – – – Principal component analysis (PCA) Independent component analysis (ICA) Fisher discriminant analysis (FDA) Random component analysis (RCA) • For fun and to show the actual gain of using different bases is relatively small • Corresponding linear methods in the spectral histogram space including – SPCA, SICA, SFDA, and SRCA May 29, 2016 Computer Vision 33 COIL Dataset May 29, 2016 Computer Vision 34 3D Recognition Results May 29, 2016 Computer Vision 35 Experimental Results - continued • To further demonstrate the effectiveness of our method for different types of images, we create a dataset of combining the texture dataset, face dataset, and COIL dataset, resulting in a dataset of 180 categories with 10160 images in total May 29, 2016 Computer Vision 36 Linear Subspaces of Spectral Representation May 29, 2016 Computer Vision 37 Experimental Results - continued • Combined dataset – continued – Not only the recognition rate is very good, but also it is very reliable and robust, as the average entropy of the p0(i|I) is 0.60 bit (The corresponding uniform distribution’s entropy is 7.49 bits) May 29, 2016 Computer Vision 38 Experimental Results - continued • Combined dataset – continued – Not only the recognition rate is very good, but also it is very reliable and robust, as the average entropy of the p0(i|I) is 0.60 bit (The corresponding uniform distribution’s entropy is 7.49 bits) Entropy=6.78bits Entropy=0.60 bit May 29, 2016 Computer Vision 39