,a),b),,,,,,,, (DNN),,,, (CNN),,.,,,,,,,,,,,,,,,,,, [], [6], [7], [], [3]., [8], [0], [7],,,, Tohoku University a) omokawa@vision.is.tohoku.ac.jp b) okatani@vision.is.tohoku.ac.jp, [3],, (DNN), DNN, [3], DNN,,, DNN,,,,,,,, DNN, DNN, (CNN)[4], [9], (i) CNN, c 03 Information Processing Society of Japan
(ii) CNN (iii) CNN., [7], [], [3] Motoyoshi [3],,,,,,,, [], [6],,,, Dror [6],, SVM,,,,,, [5], Abe [], Liu [0],,,, SIFT, Edge-slice, Hu [8] Large Margin Nearest Neighbor Liu, Timofte van Gool, [5] BIF(Basic Image Features),, [7],,,,, 3. 3., CG, Autodesk Maya., Maya. : Interior. : Nature. mental ray,, mentalray mia material, miamaterial,,, 8 8 3.., Dosch Design Interior NatureV HDR 40, 80,, Interior, Nature, DNN,,, y 0, 7, 44, 6, 88,, x, z 30, 5, 0, 5, 30,, x, z,,,, 5 5 5 5 3.. mia material, 5, (refl_gloss),,., 90 (brdf_90_degree_refl).0, 0 (brdf_0_degree_refl), 0.0.0..0,., 5, (0.,0.8), (0.8,0.8), (0.5,0.5), (0.8,0.), (0.,0.).,. c 03 Information Processing Society of Japan
3 4 5 5. ID. Interior, Nature. 3.,.,,., Interior Nature 3, 8.,, 8000( 3 5), 000( 8 5). 3 CNN.,,, softmax. CNN,,.,,,,., CNN, 50. 3.3, cuda-convnet * CNN.,,.,, C, P,, N, N.,,,,., C7, x, y, 7 7.,, 3,,, 7 7., f(x) =max(0,x) NN [9].,,,,. 3, CNN, C 7P 5 N 4 C 3P 3 N. CNN, 5,. * http://code.google.com/p/cuda-convnet/, [9]. 3.4 3.4. SIFT-BoF CNN, bag-of-features,, 3 SIFT, 7 7, SIFT 000 visual words visual words, k-means, BoF,, SVM,,,, BoF, SIFT-BoF,,,, CNN,, 3.4.,, 0 5 4,, Interior, Nature 8,, 80(= 5 8), 80 c 03 Information Processing Society of Japan 3
0.95 0.03 0.0 0.94 0.0 0.04 0.00 0.006 0.97 0.03 3 0.03 0.05 0.75 0.7 4 0.05 0.3 0.70 0.0 0.80 0.8 0.0 3 0.00 0.00 0.99 0.003 4 0.009 0.08 0.3 0.59 0.006 5 0.09 0.0 0.89 5 0.37 0.004 0.07 0.03 0.53 4 5. CNNsingle, -double and -triple,,,3 CNN. 3 4 5 (a) 0.90 0.00 0.00 0.0 3 4 5 (b) 0.97 0.00 0.03, 3, 5, 80,,,, EIZO FlexScan SX76W LCD 4. 5 0.00 0.87 0.06 0.07 3 0.0 0.006 0.85 0.3 0.003 4 0.003 0.005 0. 0.85 0.0 5 0.0 0.003 0.0 0.87 3 4 5 (c) 0.8 0.6 0.03 3 0.0 0.77 0.3 4 0.07 0.93 0.00 5 0.07 0.0 0.0 0.90 3 4 5 (d). (a). (b) CNN( ). (c) CNN( ). (d) CNN(3 ). 4. 4,,, SIFT-BoF(SVM),,,,3 CNN,,, 3 CNN 5, CNN, CNN,, C7P 3N 8, C7P 5 N4 C3P 3 N, C7P 3 N 9 C4 N 9 P3 C3 N 9 P3, CNN, SIFT-BoF(SVM),, 3 CNN, CNN,, 5,, CNN, CNN, 5, 3 4, CNN, 3, 4. 6, CNN,, [0:55], CNN,,,,,, 7 Motoyoshi [3].,,,.,,.,,,,,, Motoyoshi, [3],,,, 7,,, NN., CNN [9], c 03 Information Processing Society of Japan 4
CNN( ) CNN( ) CNN(3 ) 6. CNN,,,,. [4] CNN,,,, miamaterial,,, 5,,,,, 4 8, 5 CNN,.% 6.95% CNN,,,,, 4, 5, 5 8,,,,,., 6,,,,, CNN, 5,,, Interiror Nature, CNN 95, 8,,, CNN,,,.,, CNN,,,, [9],, 5., CNN, CNN,,, CNN, CNN,, Motoyoshi,,, Motoyoshi,, [], CNN,, [] T. Abe, T. Okatani, and K. Deguchi. Recognizing surface qualities from natural images based on learning to rank. In Proc. ICPR, pages 37-375, 0. [] B. L. Anderson and J. Kim. Image statistics do not explain the perception of gloss and lightness. journal of Vision, 9(), 009. [3] Y. Bengio, A. C. Courville, and P. Vincent. Unsuper- c 03 Information Processing Society of Japan 5
(Interior/Nature) 8 CNN 3.,,, (Interior/Nature). vised feature learning and deep learning: A review and new perspectives. CoRR, abs/06.5538, 0. [4] D. C. Ciresan, U. Meier, and J. Schmidhuber. Multicolumn deep neural networks for image classification. In Proc. CVPR, 0. [5] M. Crosier and L. D. Griffin. Using basic image features for texture classification. International Journal of Computer Vision, 88(3):447-460, 00. [6] R. O. Dror, E. H. Adelson, and A. S. Willsky. Recognition of surface reflectance properties from a single image under unknown real-world illumination. In Proc. the Workshop on Identifying Objects Across Variations in Lighting at CVPR, 00. [7] C. Hiramatsu, N. Goda, and H. Komatsu. Transformation from image-based to perceptual representation of materials along the human ventral visual pathway. NeuroImage, 57():48-494, 0. [8] D. Hu, L. Bo, and X. Ren. Toward robust material recognition for everyday objects. In Proc. BMVC, pages 48.- 48., 0. [9] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. In NIPS, 0. [0] C. Liu, L. Sharan, E. H. Adelson, and R. Rosenholtz. Exploring features in a bayesian framework for material recognition. In Proc. CVPR, 00. [] S. Lombardi and K. Nishino. Reflectance and natural illumination from a single image. In Proc. ECCV, 0. [] I. Motoyoshi. Highlight shading relationship as a cue for the perception of translucent and transparent materials. Journal of Vision, 0(9), 00. [3] I. Motoyoshi, S. Nishida, L. Sharan, and E. H. Adelson. Image statistics and the perception of surface qualities. Nature, 447(74):06-09, 007. [4] B. A. Olshausen and D. J. Field. Emergence of simplecell receptive field properties by learning a sparse code for natural images. Nature, 38(6583):607-609, 996. [5] D. Parikh and K. Grauman. Relative attributes. In Proc. ICCV, 0. [6] F. Romeiro and T. Zickler. Blind reflectrometry. In Proc. ECCV, 00. [7] R. Timofte and L. Van Gool. A training-free classification framework for textures, writers,and materials. In Proc. BMVC, pages 93.-93., 0. c 03 Information Processing Society of Japan 6