nips nips2002 nips2002-182 knowledge-graph by maker-knowledge-mining
Source: pdf
Author: William T. Freeman, Antonio Torralba
Abstract: The goal of low-level vision is to estimate an underlying scene, given an observed image. Real-world scenes (eg, albedos or shapes) can be very complex, conventionally requiring high dimensional representations which are hard to estimate and store. We propose a low-dimensional representation, called a scene recipe, that relies on the image itself to describe the complex scene configurations. Shape recipes are an example: these are the regression coefficients that predict the bandpassed shape from image data. We describe the benefits of this representation, and show two uses illustrating their properties: (1) we improve stereo shape estimates by learning shape recipes at low resolution and applying them at full resolution; (2) Shape recipes implicitly contain information about lighting and materials and we use them for material segmentation.
Reference: text
sentIndex sentText sentNum sentScore
1 We propose a low-dimensional representation, called a scene recipe, that relies on the image itself to describe the complex scene configurations. [sent-6, score-0.377]
2 Shape recipes are an example: these are the regression coefficients that predict the bandpassed shape from image data. [sent-7, score-1.312]
3 Strong priors [14] are often needed, which can give unrealistic shape reconstructions. [sent-14, score-0.547]
4 We assume that the image is always available and we describe the underlying scene in reference to the image. [sent-17, score-0.296]
5 The scene representation is a set of rules for transforming from the local image information to the desired scene quantities. [sent-18, score-0.449]
6 We call this representation a scene recipe: a simple function for transforming local image data to local scene data. [sent-19, score-0.475]
7 The computer doesn’t have to represent every curve of an intricate shape; the image does that for us, the computer just stores the rules for transforming from image to scene. [sent-20, score-0.46]
8 In this paper, we focus on reconstructing the shapes that created the observed image, deriving shape recipes. [sent-21, score-0.592]
9 The particular recipes we study here are regression coefficients for transforming (a) (b) (c) (c) Figure 1: 1-d example: The image (a) is rendered from the shape (b). [sent-22, score-1.304]
10 The shape depends on the image in a non-local way. [sent-23, score-0.762]
11 Bandpass filtering both signals allows for a local shape recipe. [sent-24, score-0.573]
12 The dotted line (which agrees closely with true solid line) in (d) shows shape reconstruction from 9-parameter linear regression (9-tap convolution) from bandpassed image, (c). [sent-25, score-0.661]
13 2 Shape Recipes The shape representation consists in describing, for a particular image, the functional relationship between image and shape. [sent-27, score-0.812]
14 To simplify the computation to obtain shape from image data, we require that the scene recipes be local: the scene structure in a region should only depend on a local neighborhood of the image. [sent-30, score-1.425]
15 1 (a) shows the intensity profile of a 1-d image arising from the shape profile shown in Fig. [sent-33, score-0.792]
16 Note that the function to recover the shape from the image cannot be local because the identical local images on the left and right sides of the surface edge correspond to different shape heights. [sent-35, score-1.452]
17 In order to obtain locality in the shape-image relationship, we need to preprocess the shape and image signals. [sent-36, score-0.762]
18 When shape and image are represented in a bandpass pyramid, within a subband, under generic rendering conditions [4], local shape changes lead to local image changes. [sent-37, score-1.68]
19 (Representing the image in a Gaussian pyramid also gives a local relationship between image and bandpassed shape, effectively subsuming the image bandpass operation into the shape recipe. [sent-38, score-1.489]
20 In this example, (d) relates to (c) by a simple shape recipe: convolution with a 9-tap filter, learned by linear regression from rendered random shape data. [sent-41, score-1.197]
21 For 2-d images, we break the image and shape into subbands using a steerable pyramid [13], an oriented multi-scale decomposition with non-aliased subbands (Fig. [sent-44, score-1.056]
22 A shape subband can be related to an image intensity subband by a function Zk = fk (Ik ) (1) where fk is a local function and Zk and Ik are the kth subbands of the steerable pyramid representation of the shape and image, respectively. [sent-46, score-2.296]
23 The simplest functional relationship between shape and image intensity is via a linear filter with a finite size impulse response: Zk ≈ rk Ik , where is convolution. [sent-47, score-0.878]
24 The convolution kernel rk (specific to each scale and orientation) transforms the image subband Ik into the shape subband Zk . [sent-48, score-1.402]
25 The recipe rk at each subband is learned by minimizing x |Zk − Ik rk |2 , regularizing rk as needed to avoid overfitting. [sent-49, score-0.562]
26 (a) Image (b) Stereo shape (c) Stereo shape (surface plot) (d) Re-rendered stereo shape Figure 2: Shape estimate from stereo. [sent-52, score-2.011]
27 (a) is one image of the stereo pair; the stereo reconstruction is depicted as (b) a range map and (c) a surface plot and (d) a re-rendering of the stereo shape. [sent-53, score-1.351]
28 The stereo shape is noisy and misses fine details. [sent-54, score-0.916]
29 We conjecture that multiscale shape recipes have various desirable properties for estimation. [sent-55, score-1.023]
30 First, they allow for a compact encoding of shape information, as much of the complexity of the shape is encoded in the image itself. [sent-56, score-1.309]
31 The recipes need only specify how to translate image into shape. [sent-57, score-0.689]
32 Secondly, regularities in how the shape recipes f k vary across scale and space provide a powerful mechanism for regularizing shape estimates. [sent-58, score-1.633]
33 Instead of regularizing shape estimates by assuming a prior of smoothness of the surface, we can assume a slow spatial variation of the functional relationship between image and shape, which should make estimating shape recipes easier. [sent-59, score-1.879]
34 Third, shape recipes implicitly encode lighting and material information, which can be used for material-based segmentation. [sent-60, score-1.223]
35 In the next two sections we discuss the properties of smoothness across scale and space and we show potential applications in improving shape estimates from stereo and in image segmentation based on material properties. [sent-61, score-1.358]
36 2 shows one image of a stereo pair and the associated shape estimated from a stereo algorithm1. [sent-63, score-1.45]
37 The shape estimate is noisy in the high frequencies (see surface plot and rerendered shape), but we assume it is accurate in the low spatial frequencies. [sent-64, score-0.686]
38 3 shows the steerable pyramid representations of the image (a) and shape (b) and the learned shape recipes (c) for each subband (linear convolution kernels that give the shape subband from the image subband). [sent-66, score-3.326]
39 4 (a) and (b) show the image and the implicit shape representation: the pyramid’s lowresolution shape and the shape recipes used over the top four scales. [sent-69, score-2.317]
40 4 (c) and (d) show explicitly the reconstructed shape implied by (a) and (b): note the high resolution details, including the fine structure visible in the bottom left corner of (d). [sent-71, score-0.597]
41 Compare with the stereo 1 We took our stereo photographs using a 3. [sent-72, score-0.688]
42 3 Megapixel Olympus Camedia C-3040 camera, with a Pentax stereo adapter. [sent-73, score-0.344]
43 We calibrated the stereo images using the point matching algorithm of Zhang [18], and rectified the stereo pair (so that epipoles are along scan lines) using the algorithm of [8], estimating disparity with the Zitnick–Kanade stereo algorithm [19]. [sent-74, score-1.064]
44 (c) Shape recipes for each subband (a) Image pyramid (b) Shape pyramid Figure 3: Learning shape recipes at each subband. [sent-77, score-1.997]
45 (a) and (b) are the steerable pyramid representations [13] of image and stereo shape. [sent-78, score-0.769]
46 The steerable pyramid isolates information according to scale (the smaller subband images represent larger spatial scales) and orientation (clockwise among subbands of one size: vertical, diagonal, horizontal, other diagonal). [sent-80, score-0.611]
47 (a) image (b) low-res shape (center, top row) and recipes (for each subband orientation) (c) recipes shape (surface plot) (d) re-rendered recipes shape Figure 4: Reconstruction from shape recipes. [sent-81, score-4.052]
48 The shape is represented by the information contained in the image (a), the low-res shape pyramid residual and the shape recipes (b) estimated at the lowest resolution. [sent-82, score-2.448]
49 The shape can be regenerated by applying the shape recipes (b) at the 4 highest resolution scales, then reconstructing from the shape pyramid. [sent-83, score-2.147]
50 (d) shows the image re-rendered under different lighting conditions than (a). [sent-84, score-0.297]
51 The reconstruction is not noisy and shows more detail than the stereo shape, Fig. [sent-85, score-0.369]
52 2, including the fine textures visible at the bottom left of the image (a) but not detected by the stereo algorithm. [sent-86, score-0.595]
53 4 Segmenting shape recipes Segmenting an image into regions of uniform color or texture is often an approximation to an underlying goal of segmenting the image into regions of uniform material. [sent-89, score-1.54]
54 Shape recipes, by describing how to transform from image to shape, implicitly encode both lighting and material properties. [sent-90, score-0.445]
55 Across unchanging lighting conditions, segmenting by shape recipes allows us to segment according to a material’s rendering properties, even overcoming changes of intensities or texture of the rendered image. [sent-91, score-1.251]
56 ) We expect shape recipes to vary smoothly over space except for abrupt boundaries at changes in material or illumination. [sent-93, score-1.176]
57 Within each subband, we can write the shape Z k (a) Shape (b) Image (c) Image-based segmentation (d) Recipe-based segmentation Figure 5: Segmentation example. [sent-94, score-0.659]
58 Based on image information alone, it is difficult to find a good segmentation into 2 groups, (c). [sent-96, score-0.271]
59 A segmentation into 2 different shape recipes naturally falls along the vertical material boundary, (d). [sent-97, score-1.207]
60 as a mixture of recipes: N p(Zk |Ik ) = p(Zk − fk,n (Ik ))pn (2) n=1 where N specifies the number of recipes needed to explain the underlying shape Z k . [sent-98, score-1.008]
61 To estimate the parameters of the mixture (shape recipes and weights), given known shape and the associated image, we use the EM algorithm [17]. [sent-100, score-1.034]
62 ) The shape recipes encode the relationship between image and shape when segmenting into 2 groups, and finds the vertical material boundary, (d). [sent-109, score-2.007]
63 These cases need to be treated specially with shape recipes. [sent-112, score-0.547]
64 6 (c) the occluding boundary in the shape only produces a smooth change in the image, Fig. [sent-114, score-0.583]
65 In that region, a shape recipe will produce an incorrect shape estimate, however, the stereo algorithm will often succeed at finding those occlusion edges. [sent-116, score-1.585]
66 On the other hand, stereo often fails to provide the shape of image regions with complex shape details, where the shape recipes succeed. [sent-117, score-2.675]
67 For the special case of revising the stereo algorithm’s output using shape recipes, we propose a statistical framework to combine both sources of information. [sent-118, score-0.891]
68 Image in full-res (a) and one steerable pyramid subband (b); stereo depth, full-res (c) and subband (d). [sent-120, score-1.068]
69 (e) shows subband of shape reconstruction using learned shape recipe. [sent-121, score-1.399]
70 Direct application of shape recipe across occlusion boundary misses the shape discontinuity. [sent-122, score-1.305]
71 Stereo algorithm catches that discontinuity, but misses other shape details. [sent-123, score-0.572]
72 Probabilistic combination of the two shape estimates (f, subband, g, surface), assuming Laplacian shape statistics, captures the desirable details of both, comparing favorably with laser scanner ground truth, (h, subband, i, surface, at slight misalignment from photos). [sent-124, score-1.136]
73 via shape recipes: p(Z|S, I) = p(S, I|Z)p(Z)/p(S, I) (3) (For notational simplicity, we omit the spatial dependency from I, S and Z. [sent-125, score-0.581]
74 ) As both stereo S and image intensity I provide strong constraints for the possible underlying shape Z, the factor p(Z) can be considered constant in the region of support of p(S, I|Z). [sent-126, score-1.15]
75 (3) can be simplified by assuming that the shapes from stereo and from shape recipes are independent. [sent-129, score-1.383]
76 Furthermore, we also assume independence between the pixels in the image and across subbands: p(S, I|Z) = p(Sk |Zk )p(Ik |Zk ) (4) k x,y Sk , Zk and Ik refer to the outputs of the subband k. [sent-130, score-0.504]
77 The terms p(Sk |Zk ) and p(Ik |Zk ) will depend on the noise models for the depth from stereo and for the shape recipes. [sent-132, score-0.915]
78 For the shape estimate from stereo we assume a Gaussian distribution for the noise. [sent-133, score-0.917]
79 At each subband and spatial location we have: 2 p(Sk |Zk ) = ps (Zk − Sk ) = 2 e−|Zk −Sk | /σs (2π)1/2 σs (5) In the case of the shape recipes, a Gaussian noise model is not adequate. [sent-134, score-0.847]
80 The distribution of the error Zk − fk (Ik ) will depend on image noise, but more importantly, on all shape and image variations that are not functionally related with each other through the recipes. [sent-135, score-1.065]
81 When trying to estimate shape using the shape recipe f k (Ik ), it fails to capture the discontinuity although it captures correctly other texture variations, Fig. [sent-140, score-1.287]
82 Therefore, Zk − fk (Ik ) will describe the distribution of occluding edges that do not produce image variations and paint edges that do not translate into shape variations. [sent-142, score-0.931]
83 The least square estimate for the shape subband Zk given both stereo and image data, is: Zk p(Sk |Zk )p(Ik |Zk )dZk (7) p(Sk |Zk )p(Ik |Zk )dZk This integral can be evaluated numerically independently at each pixel. [sent-149, score-1.398]
84 When p = 2, then the LSE estimation is a weighted linear combination of the shape from stereo and shape recipes. [sent-150, score-1.438]
85 However, with p 1 this problem is similar to the one of image denosing from wavelet decompositions [12] providing a non-linear combination of stereo and shape recipes. [sent-151, score-1.121]
86 (7) is to take from the stereo everything that cannot be explained by the recipes, and to take from the recipes the rest. [sent-153, score-0.805]
87 Whenever both stereo and shape recipes give similar estimates, we prefer the recipes because they are more accurate than the stereo information. [sent-154, score-2.157]
88 Where stereo and shape recipes differ greatly, such as at occlusions, then the shape estimate follows the stereo shape. [sent-155, score-2.269]
89 ˆ Zk = Zk p(Zk |Sk , Ik )dZk = 6 Discussion and Summary Unlike shape-from-shading algorithms [5], shape recipes are fast, local procedures for computing shape from image. [sent-156, score-1.581]
90 The approximation of linear shading [7] also assumes a local linear relationship between image and shape subbands. [sent-157, score-0.877]
91 We have proposed shape recipes as a representation that leaves the burden of describing shape details to the image. [sent-159, score-1.622]
92 Unlike many other shape representations, these are lowdimensional, and should change slowly over time, distance, and spatial scale. [sent-160, score-0.581]
93 We developed a shape estimate improver that relies on an initial estimate being accurate at low resolutions. [sent-163, score-0.599]
94 Assuming that a shape recipes change slowly over 4 octaves of spatial scale, we learned the shape recipes at low resolution and applied them at high resolution to find shape from image details not exploited by the stereo algorithm. [sent-164, score-3.275]
95 Shape recipes fold in information about both lighting and material properties and can also be used to estimate material boundaries over regions where the lighting is assumed to be constant. [sent-166, score-0.913]
96 Gilchrist and Adelson describe “atmospheres”, which are local formulas for converting image intensities to perceived lightness values [3, 1]. [sent-167, score-0.293]
97 A full description of an image in terms of a scene recipe would require both shape recipes and reflectance recipes (for computing reflectance values from image data), which also requires labelling parts of the image as being caused by shading or reflectance changes, such as [15]. [sent-169, score-2.353]
98 Using shape recipes, we find simple transformation rules that let us convert from image to shape whenever we need to, by examining the image. [sent-171, score-1.309]
99 We thank Ray Jones and Leonard McMillan for providing Cyberware scans, and Hao Zhang for code for rectification of stereo images. [sent-172, score-0.344]
100 A cooperative algorithm for stereo matching and occlusion detection. [sent-311, score-0.388]
wordName wordTfidf (topN-words)
[('shape', 0.547), ('recipes', 0.461), ('stereo', 0.344), ('zk', 0.28), ('subband', 0.266), ('image', 0.215), ('ik', 0.152), ('pyramid', 0.131), ('material', 0.122), ('recipe', 0.103), ('scene', 0.081), ('bandpassed', 0.071), ('fk', 0.07), ('sk', 0.069), ('lighting', 0.067), ('steerable', 0.061), ('surface', 0.059), ('segmentation', 0.056), ('shading', 0.055), ('rk', 0.052), ('subbands', 0.051), ('segmenting', 0.047), ('occlusion', 0.044), ('rendering', 0.04), ('convolution', 0.038), ('discontinuity', 0.037), ('lightness', 0.037), ('dzk', 0.035), ('torralba', 0.035), ('bandpass', 0.035), ('spatial', 0.034), ('relationship', 0.034), ('rendered', 0.033), ('images', 0.032), ('resolution', 0.031), ('shapes', 0.031), ('intensity', 0.03), ('transforming', 0.03), ('texture', 0.027), ('estimate', 0.026), ('ectance', 0.026), ('local', 0.026), ('reconstruction', 0.025), ('misses', 0.025), ('depth', 0.024), ('atmospheres', 0.024), ('gilchrist', 0.024), ('materials', 0.024), ('octaves', 0.024), ('paint', 0.024), ('phong', 0.024), ('zitnick', 0.024), ('regularizing', 0.023), ('recti', 0.023), ('laser', 0.023), ('across', 0.023), ('pn', 0.022), ('vertical', 0.021), ('occlusions', 0.02), ('occluding', 0.02), ('quadrants', 0.02), ('lambertian', 0.02), ('plot', 0.02), ('visible', 0.019), ('details', 0.019), ('groups', 0.019), ('boundaries', 0.019), ('scale', 0.018), ('smoothness', 0.018), ('orientation', 0.018), ('representations', 0.018), ('regression', 0.018), ('variations', 0.018), ('textures', 0.017), ('burden', 0.017), ('eg', 0.017), ('cuts', 0.016), ('boundary', 0.016), ('representation', 0.016), ('laplacian', 0.016), ('conditions', 0.015), ('properties', 0.015), ('wavelet', 0.015), ('intensities', 0.015), ('describing', 0.015), ('region', 0.014), ('regularities', 0.014), ('regions', 0.014), ('learned', 0.014), ('reconstructing', 0.014), ('changes', 0.014), ('expect', 0.013), ('translate', 0.013), ('zhang', 0.013), ('encode', 0.013), ('diagonal', 0.013), ('freeman', 0.013), ('truth', 0.013), ('implicitly', 0.013), ('edges', 0.012)]
simIndex simValue paperId paperTitle
same-paper 1 1.0 182 nips-2002-Shape Recipes: Scene Representations that Refer to the Image
Author: William T. Freeman, Antonio Torralba
Abstract: The goal of low-level vision is to estimate an underlying scene, given an observed image. Real-world scenes (eg, albedos or shapes) can be very complex, conventionally requiring high dimensional representations which are hard to estimate and store. We propose a low-dimensional representation, called a scene recipe, that relies on the image itself to describe the complex scene configurations. Shape recipes are an example: these are the regression coefficients that predict the bandpassed shape from image data. We describe the benefits of this representation, and show two uses illustrating their properties: (1) we improve stereo shape estimates by learning shape recipes at low resolution and applying them at full resolution; (2) Shape recipes implicitly contain information about lighting and materials and we use them for material segmentation.
2 0.15152723 105 nips-2002-How to Combine Color and Shape Information for 3D Object Recognition: Kernels do the Trick
Author: B. Caputo, Gy. Dorkó
Abstract: This paper presents a kernel method that allows to combine color and shape information for appearance-based object recognition. It doesn't require to define a new common representation, but use the power of kernels to combine different representations together in an effective manner. These results are achieved using results of statistical mechanics of spin glasses combined with Markov random fields via kernel functions. Experiments show an increase in recognition rate up to 5.92% with respect to conventional strategies. 1
3 0.14145054 173 nips-2002-Recovering Intrinsic Images from a Single Image
Author: Marshall F. Tappen, William T. Freeman, Edward H. Adelson
Abstract: We present an algorithm that uses multiple cues to recover shading and reflectance intrinsic images from a single image. Using both color information and a classifier trained to recognize gray-scale patterns, each image derivative is classified as being caused by shading or a change in the surface’s reflectance. Generalized Belief Propagation is then used to propagate information from areas where the correct classification is clear to areas where it is ambiguous. We also show results on real images.
4 0.13838667 10 nips-2002-A Model for Learning Variance Components of Natural Images
Author: Yan Karklin, Michael S. Lewicki
Abstract: We present a hierarchical Bayesian model for learning efficient codes of higher-order structure in natural images. The model, a non-linear generalization of independent component analysis, replaces the standard assumption of independence for the joint distribution of coefficients with a distribution that is adapted to the variance structure of the coefficients of an efficient image basis. This offers a novel description of higherorder image structure and provides a way to learn coarse-coded, sparsedistributed representations of abstract image properties such as object location, scale, and texture.
5 0.13804743 39 nips-2002-Bayesian Image Super-Resolution
Author: Michael E. Tipping, Christopher M. Bishop
Abstract: The extraction of a single high-quality image from a set of lowresolution images is an important problem which arises in fields such as remote sensing, surveillance, medical imaging and the extraction of still images from video. Typical approaches are based on the use of cross-correlation to register the images followed by the inversion of the transformation from the unknown high resolution image to the observed low resolution images, using regularization to resolve the ill-posed nature of the inversion process. In this paper we develop a Bayesian treatment of the super-resolution problem in which the likelihood function for the image registration parameters is based on a marginalization over the unknown high-resolution image. This approach allows us to estimate the unknown point spread function, and is rendered tractable through the introduction of a Gaussian process prior over images. Results indicate a significant improvement over techniques based on MAP (maximum a-posteriori) point optimization of the high resolution image and associated registration parameters. 1
7 0.11526042 183 nips-2002-Source Separation with a Sensor Array using Graphical Models and Subband Filtering
8 0.099478476 132 nips-2002-Learning to Detect Natural Image Boundaries Using Brightness and Texture
9 0.098810531 202 nips-2002-Unsupervised Color Constancy
10 0.085661247 126 nips-2002-Learning Sparse Multiscale Image Representations
11 0.084629677 74 nips-2002-Dynamic Structure Super-Resolution
12 0.068572938 57 nips-2002-Concurrent Object Recognition and Segmentation by Graph Partitioning
13 0.067114413 69 nips-2002-Discriminative Learning for Label Sequences via Boosting
14 0.066709384 44 nips-2002-Binary Tuning is Optimal for Neural Rate Coding with High Temporal Resolution
15 0.057967518 133 nips-2002-Learning to Perceive Transparency from the Statistics of Natural Scenes
16 0.052024487 127 nips-2002-Learning Sparse Topographic Representations with Products of Student-t Distributions
17 0.05195836 2 nips-2002-A Bilinear Model for Sparse Coding
18 0.050507363 87 nips-2002-Fast Transformation-Invariant Factor Analysis
19 0.049947731 206 nips-2002-Visual Development Aids the Acquisition of Motion Velocity Sensitivities
20 0.045874216 150 nips-2002-Multiple Cause Vector Quantization
topicId topicWeight
[(0, -0.115), (1, 0.017), (2, 0.001), (3, 0.253), (4, 0.004), (5, -0.056), (6, 0.174), (7, 0.005), (8, 0.011), (9, -0.051), (10, -0.011), (11, -0.04), (12, -0.01), (13, -0.027), (14, -0.016), (15, -0.024), (16, -0.053), (17, -0.043), (18, -0.037), (19, 0.004), (20, 0.051), (21, -0.012), (22, -0.052), (23, -0.029), (24, 0.075), (25, -0.019), (26, -0.021), (27, 0.017), (28, 0.028), (29, -0.076), (30, -0.014), (31, -0.078), (32, 0.093), (33, -0.042), (34, 0.077), (35, 0.053), (36, 0.01), (37, -0.001), (38, 0.029), (39, 0.106), (40, 0.122), (41, -0.056), (42, 0.018), (43, 0.162), (44, 0.104), (45, 0.015), (46, -0.092), (47, 0.186), (48, 0.161), (49, -0.027)]
simIndex simValue paperId paperTitle
same-paper 1 0.97548443 182 nips-2002-Shape Recipes: Scene Representations that Refer to the Image
Author: William T. Freeman, Antonio Torralba
Abstract: The goal of low-level vision is to estimate an underlying scene, given an observed image. Real-world scenes (eg, albedos or shapes) can be very complex, conventionally requiring high dimensional representations which are hard to estimate and store. We propose a low-dimensional representation, called a scene recipe, that relies on the image itself to describe the complex scene configurations. Shape recipes are an example: these are the regression coefficients that predict the bandpassed shape from image data. We describe the benefits of this representation, and show two uses illustrating their properties: (1) we improve stereo shape estimates by learning shape recipes at low resolution and applying them at full resolution; (2) Shape recipes implicitly contain information about lighting and materials and we use them for material segmentation.
2 0.63440013 173 nips-2002-Recovering Intrinsic Images from a Single Image
Author: Marshall F. Tappen, William T. Freeman, Edward H. Adelson
Abstract: We present an algorithm that uses multiple cues to recover shading and reflectance intrinsic images from a single image. Using both color information and a classifier trained to recognize gray-scale patterns, each image derivative is classified as being caused by shading or a change in the surface’s reflectance. Generalized Belief Propagation is then used to propagate information from areas where the correct classification is clear to areas where it is ambiguous. We also show results on real images.
3 0.58124435 202 nips-2002-Unsupervised Color Constancy
Author: Kinh Tieu, Erik G. Miller
Abstract: In [1] we introduced a linear statistical model of joint color changes in images due to variation in lighting and certain non-geometric camera parameters. We did this by measuring the mappings of colors in one image of a scene to colors in another image of the same scene under different lighting conditions. Here we increase the flexibility of this color flow model by allowing flow coefficients to vary according to a low order polynomial over the image. This allows us to better fit smoothly varying lighting conditions as well as curved surfaces without endowing our model with too much capacity. We show results on image matching and shadow removal and detection.
4 0.56328225 39 nips-2002-Bayesian Image Super-Resolution
Author: Michael E. Tipping, Christopher M. Bishop
Abstract: The extraction of a single high-quality image from a set of lowresolution images is an important problem which arises in fields such as remote sensing, surveillance, medical imaging and the extraction of still images from video. Typical approaches are based on the use of cross-correlation to register the images followed by the inversion of the transformation from the unknown high resolution image to the observed low resolution images, using regularization to resolve the ill-posed nature of the inversion process. In this paper we develop a Bayesian treatment of the super-resolution problem in which the likelihood function for the image registration parameters is based on a marginalization over the unknown high-resolution image. This approach allows us to estimate the unknown point spread function, and is rendered tractable through the introduction of a Gaussian process prior over images. Results indicate a significant improvement over techniques based on MAP (maximum a-posteriori) point optimization of the high resolution image and associated registration parameters. 1
5 0.50540179 105 nips-2002-How to Combine Color and Shape Information for 3D Object Recognition: Kernels do the Trick
Author: B. Caputo, Gy. Dorkó
Abstract: This paper presents a kernel method that allows to combine color and shape information for appearance-based object recognition. It doesn't require to define a new common representation, but use the power of kernels to combine different representations together in an effective manner. These results are achieved using results of statistical mechanics of spin glasses combined with Markov random fields via kernel functions. Experiments show an increase in recognition rate up to 5.92% with respect to conventional strategies. 1
6 0.4742097 74 nips-2002-Dynamic Structure Super-Resolution
7 0.37970009 132 nips-2002-Learning to Detect Natural Image Boundaries Using Brightness and Texture
8 0.36434782 183 nips-2002-Source Separation with a Sensor Array using Graphical Models and Subband Filtering
9 0.36325306 10 nips-2002-A Model for Learning Variance Components of Natural Images
10 0.35471871 150 nips-2002-Multiple Cause Vector Quantization
12 0.3305907 126 nips-2002-Learning Sparse Multiscale Image Representations
13 0.3190504 47 nips-2002-Branching Law for Axons
14 0.31876951 193 nips-2002-Temporal Coherence, Natural Image Sequences, and the Visual Cortex
15 0.31626007 133 nips-2002-Learning to Perceive Transparency from the Statistics of Natural Scenes
16 0.28119403 57 nips-2002-Concurrent Object Recognition and Segmentation by Graph Partitioning
17 0.27052647 2 nips-2002-A Bilinear Model for Sparse Coding
18 0.26094267 127 nips-2002-Learning Sparse Topographic Representations with Products of Student-t Distributions
19 0.23802872 117 nips-2002-Intrinsic Dimension Estimation Using Packing Numbers
20 0.23446552 44 nips-2002-Binary Tuning is Optimal for Neural Rate Coding with High Temporal Resolution
topicId topicWeight
[(1, 0.013), (3, 0.013), (11, 0.019), (23, 0.017), (42, 0.04), (45, 0.017), (54, 0.088), (55, 0.021), (61, 0.382), (64, 0.011), (67, 0.014), (68, 0.018), (74, 0.141), (92, 0.025), (98, 0.068)]
simIndex simValue paperId paperTitle
same-paper 1 0.83330411 182 nips-2002-Shape Recipes: Scene Representations that Refer to the Image
Author: William T. Freeman, Antonio Torralba
Abstract: The goal of low-level vision is to estimate an underlying scene, given an observed image. Real-world scenes (eg, albedos or shapes) can be very complex, conventionally requiring high dimensional representations which are hard to estimate and store. We propose a low-dimensional representation, called a scene recipe, that relies on the image itself to describe the complex scene configurations. Shape recipes are an example: these are the regression coefficients that predict the bandpassed shape from image data. We describe the benefits of this representation, and show two uses illustrating their properties: (1) we improve stereo shape estimates by learning shape recipes at low resolution and applying them at full resolution; (2) Shape recipes implicitly contain information about lighting and materials and we use them for material segmentation.
2 0.6242153 179 nips-2002-Scaling of Probability-Based Optimization Algorithms
Author: J. L. Shapiro
Abstract: Population-based Incremental Learning is shown require very sensitive scaling of its learning rate. The learning rate must scale with the system size in a problem-dependent way. This is shown in two problems: the needle-in-a haystack, in which the learning rate must vanish exponentially in the system size, and in a smooth function in which the learning rate must vanish like the square root of the system size. Two methods are proposed for removing this sensitivity. A learning dynamics which obeys detailed balance is shown to give consistent performance over the entire range of learning rates. An analog of mutation is shown to require a learning rate which scales as the inverse system size, but is problem independent. 1
3 0.43686074 162 nips-2002-Parametric Mixture Models for Multi-Labeled Text
Author: Naonori Ueda, Kazumi Saito
Abstract: We propose probabilistic generative models, called parametric mixture models (PMMs), for multiclass, multi-labeled text categorization problem. Conventionally, the binary classification approach has been employed, in which whether or not text belongs to a category is judged by the binary classifier for every category. In contrast, our approach can simultaneously detect multiple categories of text using PMMs. We derive efficient learning and prediction algorithms for PMMs. We also empirically show that our method could significantly outperform the conventional binary methods when applied to multi-labeled text categorization using real World Wide Web pages. 1
4 0.42615581 90 nips-2002-Feature Selection in Mixture-Based Clustering
Author: Martin H. Law, Anil K. Jain, Mário Figueiredo
Abstract: There exist many approaches to clustering, but the important issue of feature selection, i.e., selecting the data attributes that are relevant for clustering, is rarely addressed. Feature selection for clustering is difficult due to the absence of class labels. We propose two approaches to feature selection in the context of Gaussian mixture-based clustering. In the first one, instead of making hard selections, we estimate feature saliencies. An expectation-maximization (EM) algorithm is derived for this task. The second approach extends Koller and Sahami’s mutual-informationbased feature relevance criterion to the unsupervised case. Feature selection is then carried out by a backward search scheme. This scheme can be classified as a “wrapper”, since it wraps mixture estimation in an outer layer that performs feature selection. Experimental results on synthetic and real data show that both methods have promising performance.
5 0.42545739 78 nips-2002-Efficient Learning Equilibrium
Author: Ronen I. Brafman, Moshe Tennenholtz
Abstract: We introduce efficient learning equilibrium (ELE), a normative approach to learning in non cooperative settings. In ELE, the learning algorithms themselves are required to be in equilibrium. In addition, the learning algorithms arrive at a desired value after polynomial time, and deviations from a prescribed ELE become irrational after polynomial time. We prove the existence of an ELE in the perfect monitoring setting, where the desired value is the expected payoff in a Nash equilibrium. We also show that an ELE does not always exist in the imperfect monitoring case. Yet, it exists in the special case of common-interest games. Finally, we extend our results to general stochastic games. 1
6 0.42374471 175 nips-2002-Reinforcement Learning to Play an Optimal Nash Equilibrium in Team Markov Games
8 0.42102876 114 nips-2002-Information Regularization with Partially Labeled Data
9 0.41833931 39 nips-2002-Bayesian Image Super-Resolution
10 0.41501191 132 nips-2002-Learning to Detect Natural Image Boundaries Using Brightness and Texture
11 0.41499087 173 nips-2002-Recovering Intrinsic Images from a Single Image
12 0.4136802 89 nips-2002-Feature Selection by Maximum Marginal Diversity
13 0.41356325 74 nips-2002-Dynamic Structure Super-Resolution
14 0.41136461 152 nips-2002-Nash Propagation for Loopy Graphical Games
15 0.41109329 122 nips-2002-Learning About Multiple Objects in Images: Factorial Learning without Factorial Search
16 0.40903169 57 nips-2002-Concurrent Object Recognition and Segmentation by Graph Partitioning
17 0.40897834 31 nips-2002-Application of Variational Bayesian Approach to Speech Recognition
18 0.4076021 2 nips-2002-A Bilinear Model for Sparse Coding
19 0.40312684 124 nips-2002-Learning Graphical Models with Mercer Kernels
20 0.40258858 157 nips-2002-On the Dirichlet Prior and Bayesian Regularization