iccv iccv2013 iccv2013-47 knowledge-graph by maker-knowledge-mining
Source: pdf
Author: Samuel Schulter, Christian Leistner, Paul Wohlhart, Peter M. Roth, Horst Bischof
Abstract: We present Alternating Regression Forests (ARFs), a novel regression algorithm that learns a Random Forest by optimizing a global loss function over all trees. This interrelates the information of single trees during the training phase and results in more accurate predictions. ARFs can minimize any differentiable regression loss without sacrificing the appealing properties of Random Forests, like low computational complexity during both, training and testing. Inspired by recent developments for classification [19], we derive a new algorithm capable of dealing with different regression loss functions, discuss its properties and investigate the relations to other methods like Boosted Trees. We evaluate ARFs on standard machine learning benchmarks, where we observe better generalization power compared to both standard Random Forests and Boosted Trees. Moreover, we apply the proposed regressor to two computer vision applications: object detection and head pose estimation from depth images. ARFs outperform the Random Forest baselines in both tasks, illustrating the importance of optimizing a common loss function for all trees.
Reference: text
sentIndex sentText sentNum sentScore
1 cg , , , Abstract We present Alternating Regression Forests (ARFs), a novel regression algorithm that learns a Random Forest by optimizing a global loss function over all trees. [sent-4, score-0.386]
2 This interrelates the information of single trees during the training phase and results in more accurate predictions. [sent-5, score-0.156]
3 ARFs can minimize any differentiable regression loss without sacrificing the appealing properties of Random Forests, like low computational complexity during both, training and testing. [sent-6, score-0.45]
4 Inspired by recent developments for classification [19], we derive a new algorithm capable of dealing with different regression loss functions, discuss its properties and investigate the relations to other methods like Boosted Trees. [sent-7, score-0.355]
5 Moreover, we apply the proposed regressor to two computer vision applications: object detection and head pose estimation from depth images. [sent-9, score-0.281]
6 During training, each tree of the forest is grown independently by recursively splitting the labeled training data until some stopping criteria are fulfilled, in order to disambiguate the uncertainty of the predictions. [sent-18, score-0.389]
7 com of the trees results in a strong and highly non-linear predictor, applicable to many different tasks like classification, regression or density estimation [5]. [sent-22, score-0.376]
8 Recently, RFs have also been extensively used in computer vision to solve different regression tasks. [sent-28, score-0.225]
9 In this paper, we propose a novel Random Forest training procedure named Alternating Regression Forests (ARFs), which, inspired by [19], globally optimizes differentiable regression loss functions. [sent-31, score-0.443]
10 In particular, we formulate the training of Random Regression Forests as a general risk minimization problem and model the whole forest as a stage-wise classifier, akin to Gradient Boosting [12] and ADFs [19]. [sent-32, score-0.207]
11 The iterative training procedure alternates between growing the forest by one level and evaluating the global loss for all training samples. [sent-33, score-0.401]
12 Additionally, we discuss the properties of ARFs, the influence of different choices of the regression loss function and give some insights about the relations to other approaches like Boosted Trees or standard Random Regression Forests. [sent-42, score-0.377]
13 First, we show regression performance on several standard machine learning benchmarks and analyze the relevant parameters. [sent-45, score-0.278]
14 Second, we extend Hough Forests [9] for object detection by applying the ARF principle to the regression split nodes and show improved results on three test sets. [sent-46, score-0.329]
15 Finally, we integrate ARFs into the human head pose estimation system of Fanelli et al. [sent-47, score-0.152]
16 Related Work Recently, Random Forests (RFs) have been increasingly employed for difficult regression tasks in several computer vision applications, enabled by the non-parametric and highly non-linear structure of RFs. [sent-50, score-0.245]
17 The prediction of the RF is conditioned on an estimated head pose and gives state-of-the-art results on standard benchmarks. [sent-54, score-0.155]
18 Another task is human head pose estimation [7] from single depth images, where a Random Regression Forest is trained to estimate the position of the head (e. [sent-55, score-0.376]
19 Similar approaches have also been used for human pose estimation [10], where several body joint locations are regressed from single depth images. [sent-58, score-0.182]
20 RFs have also been employed in joint classification and regression tasks, where the objective function is formulated jointly for both tasks. [sent-59, score-0.225]
21 Hough Forests (HFs) for object detection [9] or head detection and simultaneous pose estimation [8] are just two examples. [sent-60, score-0.152]
22 4), we show that integrating our proposed regression algorithm into two of these applications gives better performance. [sent-62, score-0.225]
23 The proposed Regression Tree Field builds on a Gaussian Random Field and its parameters are trained from image data with regression trees. [sent-64, score-0.244]
24 The paper also shows how the regression trees can be optimally trained for minimizing the Random Field energy. [sent-66, score-0.356]
25 In contrast to single tree optimization, we present a regression algorithm based on an ensemble of trees and show how this ensemble can be optimized with a global loss function. [sent-67, score-0.563]
26 Boosted Trees for regression [12] are also related to our proposed regressor. [sent-69, score-0.225]
27 Alternating Regression Forests To derive Alternating Regression Forests (ARFs), we first briefly review standard Random Regression Forests and show how a global loss function can be integrated into Random Forests (RFs) for classification [19]. [sent-76, score-0.157]
28 Finally, we discuss the properties of the new algorithm, the employed loss functions and point out the relations between ARFs, RFs and Gradient Boosting [12]. [sent-81, score-0.161]
29 Random Forests for Regression In general, for regression we are given labeled training samples {xi, yi}iN=1, where xi ∈ X = RM and yi ∈ Y = sRaKm,p sampled fro}m a joint probability d Ristribution q∈(x Y, Yy =). [sent-84, score-0.312]
30 This mapping is learned by an ensemble of binary decision trees {Tt}tT=1 (T being the naunm enbseerm mofb etre oefs biinn tahrey deencsiesmiobnle t)r,e eeasc {hT tr}ained on a subset of the training data (c. [sent-87, score-0.232]
31 A single decision tree Tt recursively splits the given training data into two partitTions, such that the uncertainty of the target variables in the resulting subsets is minimized. [sent-90, score-0.214]
32 In particular, each node in a tree randomly samples a set of splitting functions φ(x), each separating the data into two disjoint subsets, L and R, respectively. [sent-91, score-0.227]
33 All splitting functions are then evaluated by measuring the information gain I = H(L∪R)−|L||L +| |R|H(L)−|L||R +| |R|H(R) , (1) where H(·) is the entropy over the target labels and | · | denwohteesre eth He (s·i)ze is so tfh a s eent. [sent-92, score-0.16]
34 For regression tasks, the differential entropy h(q) =Y? [sent-94, score-0.246]
35 , a maximum tree depth or a minimum number of samples left in the splitting node. [sent-100, score-0.246]
36 If one of these criteria is fulfilled, a leaf node is created by estimating a density model p(y) from all samples falling in this leaf, in order to predict the target value. [sent-101, score-0.2]
37 While this allows for easy parallelization and thus leads to low computational costs, the learning procedure is not globally controlled by an appropriate loss function. [sent-107, score-0.157]
38 The trees are trained breadth-first up to depth Dmax and each level of depth d corresponds to a single stage. [sent-109, score-0.323]
39 he T hloenss, for each training sample can be calculated and exploited to optimize a global loss function over the whole Random Forest in the next stage d. [sent-116, score-0.206]
40 Thus, ADFs optimize a global loss over all trees by keeping a weight distribution over the training samples, which gets updated in each stage d according to the given loss function and the current state of the classifier. [sent-121, score-0.43]
41 Contrary to [17, 15], where only nodes within a single tree are entangled, ADFs entangle all trees in the forest, as each local node split depends on the output of all other trees. [sent-123, score-0.317]
42 In the following, we exploit these ideas to develop Alternating Regression Forests, which can optimize any differentiable, global regression loss. [sent-124, score-0.248]
43 Training Alternating Regression Forest We now formulate Alternating Regression Forests as a stage-wise risk minimization problem for regression tasks. [sent-127, score-0.243]
44 i,yi}l(yi;FDmax(xi;Θ) , (4) where l(·) is a differentiable loss function and FDmax = fd(x, Θd) denotes the random forest with a tree d? [sent-129, score-0.37]
45 To keep the notation uncluttered, Θ denotes all splitting functions of the corresponding forest F(·). [sent-137, score-0.229]
46 Please note that training depth d of the forest corresponds to transforming the leaf nodes in Fd−1 (x) into splitting nodes and creating new leaf nodes in depth d, which make predictions about the “pseudo targets” −gdi. [sent-162, score-0.806]
47 , the new split node in depth d − 1, to the current target distributions sppclh(ity n)o odfe t ihne new hc dhi −ld 1n,o tdoe tsh ien depth td t. [sent-166, score-0.312]
48 A given test sample x is routed through all trees Tt and thus genivdesn up itn s aTm lpelaef xno idse ros,u eteadch th storing litls reesetism Tated target variable distribution p(y). [sent-171, score-0.186]
49 Discussion In the following, we specify the loss functions used in our implementation and briefly discuss the properties of the algorithm compared to related approaches. [sent-181, score-0.161]
50 In general, any differentiable loss function can be integrated into ARFs, however, we use three of the most common regression losses, the Squared, the Absolute, and the Huber loss, well known from robust statistics [12]. [sent-182, score-0.38]
51 It behaves like a Squared loss for residuals below δ and like an Absolute loss for residuals larger than δ. [sent-187, score-0.298]
52 , the number of splitting functions is the same, for ARFs, this space increases over time as the number of split functions to be optimized increases in each iteration. [sent-192, score-0.178]
53 Contrary, in ARFs, a weak learner corresponds to the splitting functions in a single depth d. [sent-197, score-0.279]
54 This implies that for d > 0 the training samples are conditioned on the structure of the trees up to depth d − 1, which eases tohne thaesk s tfruorc tthuree ew oefa thk ele tarreneesr u ipn tthoe d ceputrhren dt − −ite 1r,at wihonic. [sent-198, score-0.272]
55 In contrast, ARFs regard each depth of the forest as a single weak leaner, i. [sent-204, score-0.263]
56 We first build a random train-test split (unless an explicit split is given) with the above defined ratio and then, for each split, we train and test all methods 4 times in order to further decrease statistical uncertainties due to the random tree growing schemes. [sent-218, score-0.247]
57 The parameters of all trees for the different methods (RFs, BTs, and ARFs) are set equally: we used 50 trees with a maximum depth of 15, however, tree growing√ √also stops if the sample size in a node is below 10. [sent-221, score-0.421]
58 53 10203 4ABR0 TF0S qr50 # trees # trees (a) Dmax = 5 (b) Dmax = 15 Figure 2: Parameter evaluation of RFs, BTs and ARFs for the experiment conducted on autompg. [sent-241, score-0.224]
59 We vary the number of trees for two choices of the maximum tree depth Dmax. [sent-242, score-0.263]
60 We make three observations: First, a larger amount of weak learners is important for BTs (both plots), which, however, also implicates a longer training time compared to RFs and ARFs, as no parallelization is possible. [sent-250, score-0.191]
61 Second, BTs can handle shallow trees as weak learners much better than RFs or ARFs (see Fig. [sent-251, score-0.233]
62 Please note that these values correspond to 50 weak learners and would even be more distinctive if the number of weak learners is increased. [sent-259, score-0.242]
63 We first give a brief review of HFs and describe our modifications to the regression nodes, before we present our results on three popular object detection benchmarks. [sent-263, score-0.225]
64 During training, a Random Forest is built to disambiguate both the class uncertainty of all patches and the regression variance of foreground patches. [sent-267, score-0.295]
65 003 5 3 Table 1: Machine learning results for the compared methods (RFs, BTs, ARFs) with different loss functions (Absolute, Squared, Huber) for BTs and ARFs on standard regression benchmarks. [sent-791, score-0.39]
66 Each split node is randomly assigned to be either a classification or a regression node. [sent-794, score-0.312]
67 Classification nodes are simi- larly designed as in standard RFs and extended to optimize a global loss in [19]. [sent-795, score-0.22]
68 In the following, we further extend HFs in the regression nodes. [sent-796, score-0.225]
69 Regression nodes in HFs follow the simple ReductionIn-Variance approach [9], which measures the uncertainty of the node predictions as H(S) =|1S|i? [sent-797, score-0.156]
70 The trees are grown depth by depth and after each iteration the “pseudo targets” are calculated via the given loss and the current prediction of the forest. [sent-804, score-0.436]
71 HFs stop growing trees if either a maximum depth is reached or less than 20 samples are available in a node. [sent-805, score-0.264]
72 Then, leaf nodes are created that store a class probability, i. [sent-806, score-0.165]
73 While standard HFs store all offset vectors di reaching a leaf node, we follow a different approach [10], where only modes of the distribution of offset vectors are stored. [sent-809, score-0.214]
74 As we already calculate the mean of the offset vectors for the residuals during tree growing, we simply stick with these estimates for the final offset vectors in the leaf nodes. [sent-811, score-0.281]
75 We increase the maximum tree depth to 30, in order to handle occurring multi modal distributions. [sent-812, score-0.151]
76 In this experiment, we directly evaluate the influence of the global loss in the regression nodes of HFs. [sent-818, score-0.423]
77 Further, we also include the influence of a global loss for the classification nodes [19] (ADF). [sent-820, score-0.198]
78 For a fair comparison, we endow all methods with 10 trees, each having a maximum depth of 30, and 20000 random tests per node. [sent-821, score-0.155]
79 For ARFs, we use the Squared loss throughout all experiments, as this loss consistently gave the best results. [sent-825, score-0.224]
80 However, optimizing a global regression loss during the training procedure, i. [sent-830, score-0.43]
81 Application II: Head Pose Estimation Our second computer vision application for ARFs is human head pose estimation from depth images, where we follow the experimental setup of Fanelli et al. [sent-840, score-0.248]
82 The goal is to train ajoint classification and regression forest on patches from labeled depth images, in order to detect the head and to estimate the pose from unseen depth images. [sent-842, score-0.701]
83 , being a face or non-face patch) and only positive patches store target regression values (i. [sent-845, score-0.308]
84 , the head center position relative to the patch in 3D coordinates and the pose in Euler angles). [sent-847, score-0.152]
85 Thus, the Random Forest is very similar to standard HFs [9] with a few exceptions, like a different form of the splitting functions φ(x) (Haar-like features), a larger patch size of 100px, or a slightly different entropy measure H(·). [sent-848, score-0.149]
86 Each leaf node stores a foreground probability pfg, the target prediction p(y) (mean of all training sample targets), and the variance σ2 of those targets. [sent-850, score-0.221]
87 During testing, patches from the depth image are extracted on a regular grid and routed to the corresponding leaf nodes in all trees. [sent-854, score-0.308]
88 Each leaf node having a foreground probability pfg = 1and a variance σ2 < σm2ax are selected 423 to vote for a head center and pose [8]. [sent-855, score-0.318]
89 Groundtruth is given as the head center position in 3D and the pose in Euler angles for each frame. [sent-860, score-0.152]
90 Like in the previous experiment, we compare HFs, modified for this regression task [8], ADFs, and ARFs. [sent-861, score-0.225]
91 Results: Following [8], we also present our results as the percentage of correctly predicted frames over different success thresholds for both, position and pose regression of the head, see Fig. [sent-865, score-0.287]
92 Nevertheless, we can see from the plots that both approaches optimizing a global loss (ADFs and ARFs) consistently improve over the HFs baseline. [sent-870, score-0.161]
93 Although this is a regression task, we can observe that ADFs (classification nodes) significantly improve over HFs. [sent-871, score-0.225]
94 38 Table 2: Raw regression errors (mean and standard deviation) of HF [8], ARF, ADF [19] and ARF* in mm for X, Y, Z, and Position and in degree for Yaw, Pitch, Roll and Angle. [sent-1008, score-0.247]
95 Acuray%170869 0 Posit15nero2th0reso2ld5(HmA RFD )*30%ycaruA109876 0 1Angle15roth2e0sold(25eA HgrRFDe Fs*)30 (a) (b) Figure 4: Frame accuracy of the competing methods (HF, ADF, ARF and ARF*) for different success thresholds of (a) the head position in mm and (b) the head pose in degree. [sent-1009, score-0.242]
96 To get a better insight in the accuracies of all methods, we also give the raw errors for all 6 variables and the aggregated head position and head pose (angle) errors in Tab. [sent-1010, score-0.242]
97 Again, we observe that all methods optimizing a global loss consistently outperform the baseline [8] in this task. [sent-1012, score-0.161]
98 Conclusion We presented Alternating Regression Forests, a novel Random Forest training procedure for regression tasks, which, in contrast to standard Random Regression Forests, optimizes any differentiable global loss function without sacrificing the computational benefits of Random Forests. [sent-1014, score-0.514]
99 Furthermore, we also integrated our ideas into two computer vision applications (object detection with Hough Forests and pose estimation from depth images). [sent-1017, score-0.158]
100 In both cases, ARFs could beat the baselines, illustrating the benefits of optimizing a global loss during training. [sent-1018, score-0.161]
wordName wordTfidf (topN-words)
[('arfs', 0.628), ('rfs', 0.339), ('forests', 0.241), ('adfs', 0.232), ('regression', 0.225), ('bts', 0.191), ('hfs', 0.191), ('arf', 0.169), ('forest', 0.123), ('trees', 0.112), ('loss', 0.112), ('depth', 0.096), ('head', 0.09), ('alternating', 0.086), ('boosting', 0.084), ('leaf', 0.08), ('fd', 0.079), ('learners', 0.077), ('splitting', 0.075), ('nodes', 0.063), ('dmax', 0.06), ('decision', 0.058), ('hough', 0.057), ('pseudo', 0.055), ('tree', 0.055), ('adf', 0.048), ('fanelli', 0.047), ('node', 0.046), ('offset', 0.045), ('training', 0.044), ('targets', 0.044), ('weak', 0.044), ('differentiable', 0.043), ('pose', 0.043), ('pfg', 0.041), ('routed', 0.041), ('split', 0.041), ('huber', 0.04), ('boosted', 0.038), ('residuals', 0.037), ('random', 0.037), ('rmse', 0.037), ('growing', 0.036), ('learner', 0.033), ('target', 0.033), ('regressor', 0.033), ('squared', 0.032), ('benchmarks', 0.031), ('functions', 0.031), ('rf', 0.031), ('hf', 0.03), ('wohlhart', 0.029), ('patches', 0.028), ('stopping', 0.027), ('autompg', 0.027), ('fdmax', 0.027), ('pchild', 0.027), ('schulter', 0.027), ('stage', 0.027), ('euler', 0.027), ('parallelization', 0.026), ('sacrificing', 0.026), ('optimizing', 0.026), ('gall', 0.025), ('gdi', 0.024), ('regressed', 0.024), ('costs', 0.024), ('uncertainty', 0.024), ('xi', 0.023), ('global', 0.023), ('leistner', 0.023), ('predictions', 0.023), ('standard', 0.022), ('endow', 0.022), ('entangled', 0.022), ('ppa', 0.022), ('store', 0.022), ('akin', 0.022), ('criteria', 0.021), ('fulfilled', 0.021), ('jancsary', 0.021), ('entropy', 0.021), ('grown', 0.02), ('tasks', 0.02), ('gradient', 0.02), ('samples', 0.02), ('trained', 0.019), ('fiducial', 0.019), ('votes', 0.019), ('position', 0.019), ('roth', 0.019), ('procedure', 0.019), ('classifier', 0.019), ('estimation', 0.019), ('calculate', 0.019), ('pitch', 0.019), ('ensemble', 0.018), ('risk', 0.018), ('discuss', 0.018), ('foreground', 0.018)]
simIndex simValue paperId paperTitle
same-paper 1 1.0000006 47 iccv-2013-Alternating Regression Forests for Object Detection and Pose Estimation
Author: Samuel Schulter, Christian Leistner, Paul Wohlhart, Peter M. Roth, Horst Bischof
Abstract: We present Alternating Regression Forests (ARFs), a novel regression algorithm that learns a Random Forest by optimizing a global loss function over all trees. This interrelates the information of single trees during the training phase and results in more accurate predictions. ARFs can minimize any differentiable regression loss without sacrificing the appealing properties of Random Forests, like low computational complexity during both, training and testing. Inspired by recent developments for classification [19], we derive a new algorithm capable of dealing with different regression loss functions, discuss its properties and investigate the relations to other methods like Boosted Trees. We evaluate ARFs on standard machine learning benchmarks, where we observe better generalization power compared to both standard Random Forests and Boosted Trees. Moreover, we apply the proposed regressor to two computer vision applications: object detection and head pose estimation from depth images. ARFs outperform the Random Forest baselines in both tasks, illustrating the importance of optimizing a common loss function for all trees.
2 0.178784 404 iccv-2013-Structured Forests for Fast Edge Detection
Author: Piotr Dollár, C. Lawrence Zitnick
Abstract: Edge detection is a critical component of many vision systems, including object detectors and image segmentation algorithms. Patches of edges exhibit well-known forms of local structure, such as straight lines or T-junctions. In this paper we take advantage of the structure present in local image patches to learn both an accurate and computationally efficient edge detector. We formulate the problem of predicting local edge masks in a structured learning framework applied to random decision forests. Our novel approach to learning decision trees robustly maps the structured labels to a discrete space on which standard information gain measures may be evaluated. The result is an approach that obtains realtime performance that is orders of magnitude faster than many competing state-of-the-art approaches, while also achieving state-of-the-art edge detection results on the BSDS500 Segmentation dataset and NYU Depth dataset. Finally, we show the potential of our approach as a general purpose edge detector by showing our learned edge models generalize well across datasets.
3 0.13001858 336 iccv-2013-Random Forests of Local Experts for Pedestrian Detection
Author: Javier Marín, David Vázquez, Antonio M. López, Jaume Amores, Bastian Leibe
Abstract: Pedestrian detection is one of the most challenging tasks in computer vision, and has received a lot of attention in the last years. Recently, some authors have shown the advantages of using combinations of part/patch-based detectors in order to cope with the large variability of poses and the existence of partial occlusions. In this paper, we propose a pedestrian detection method that efficiently combines multiple local experts by means of a Random Forest ensemble. The proposed method works with rich block-based representations such as HOG and LBP, in such a way that the same features are reused by the multiple local experts, so that no extra computational cost is needed with respect to a holistic method. Furthermore, we demonstrate how to integrate the proposed approach with a cascaded architecture in order to achieve not only high accuracy but also an acceptable efficiency. In particular, the resulting detector operates at five frames per second using a laptop machine. We tested the proposed method with well-known challenging datasets such as Caltech, ETH, Daimler, and INRIA. The method proposed in this work consistently ranks among the top performers in all the datasets, being either the best method or having a small difference with the best one.
4 0.1267741 391 iccv-2013-Sieving Regression Forest Votes for Facial Feature Detection in the Wild
Author: Heng Yang, Ioannis Patras
Abstract: In this paper we propose a method for the localization of multiple facial features on challenging face images. In the regression forests (RF) framework, observations (patches) that are extracted at several image locations cast votes for the localization of several facial features. In order to filter out votes that are not relevant, we pass them through two types of sieves, that are organised in a cascade, and which enforce geometric constraints. The first sieve filters out votes that are not consistent with a hypothesis for the location of the face center. Several sieves of the second type, one associated with each individual facial point, filter out distant votes. We propose a method that adjusts onthe-fly the proximity threshold of each second type sieve by applying a classifier which, based on middle-level features extracted from voting maps for the facial feature in question, makes a sequence of decisions on whether the threshold should be reduced or not. We validate our proposed method on two challenging datasets with images collected from the Internet in which we obtain state of the art results without resorting to explicit facial shape models. We also show the benefits of our method for proximity threshold adjustment especially on ’difficult’ face images.
5 0.11821365 448 iccv-2013-Weakly Supervised Learning of Image Partitioning Using Decision Trees with Structured Split Criteria
Author: Christoph Straehle, Ullrich Koethe, Fred A. Hamprecht
Abstract: We propose a scheme that allows to partition an image into a previously unknown number of segments, using only minimal supervision in terms of a few must-link and cannotlink annotations. We make no use of regional data terms, learning instead what constitutes a likely boundary between segments. Since boundaries are only implicitly specified through cannot-link constraints, this is a hard and nonconvex latent variable problem. We address this problem in a greedy fashion using a randomized decision tree on features associated with interpixel edges. We use a structured purity criterion during tree construction and also show how a backtracking strategy can be used to prevent the greedy search from ending up in poor local optima. The proposed strategy is compared with prior art on natural images.
6 0.11768968 133 iccv-2013-Efficient Hand Pose Estimation from a Single Depth Image
7 0.11547909 340 iccv-2013-Real-Time Articulated Hand Pose Estimation Using Semi-supervised Transductive Regression Forests
8 0.098899387 437 iccv-2013-Unsupervised Random Forest Manifold Alignment for Lipreading
9 0.098031603 132 iccv-2013-Efficient 3D Scene Labeling Using Fields of Trees
10 0.093897253 352 iccv-2013-Revisiting Example Dependent Cost-Sensitive Learning with Decision Trees
12 0.080754578 338 iccv-2013-Randomized Ensemble Tracking
13 0.072902553 273 iccv-2013-Monocular Image 3D Human Pose Estimation under Self-Occlusion
14 0.072455242 318 iccv-2013-PixelTrack: A Fast Adaptive Algorithm for Tracking Non-rigid Objects
15 0.066000842 165 iccv-2013-Find the Best Path: An Efficient and Accurate Classifier for Image Hierarchies
16 0.063120656 5 iccv-2013-A Color Constancy Model with Double-Opponency Mechanisms
17 0.061210454 190 iccv-2013-Handling Occlusions with Franken-Classifiers
18 0.060742404 81 iccv-2013-Combining the Right Features for Complex Event Recognition
19 0.059664574 209 iccv-2013-Image Guided Depth Upsampling Using Anisotropic Total Generalized Variation
20 0.058837399 382 iccv-2013-Semi-dense Visual Odometry for a Monocular Camera
topicId topicWeight
[(0, 0.14), (1, -0.011), (2, -0.029), (3, -0.03), (4, 0.045), (5, -0.036), (6, -0.006), (7, 0.028), (8, -0.044), (9, -0.007), (10, -0.051), (11, -0.023), (12, -0.039), (13, -0.005), (14, 0.08), (15, 0.056), (16, -0.048), (17, -0.142), (18, -0.007), (19, 0.097), (20, -0.054), (21, 0.066), (22, -0.04), (23, 0.115), (24, -0.058), (25, -0.02), (26, 0.016), (27, 0.043), (28, 0.021), (29, -0.111), (30, 0.066), (31, 0.057), (32, -0.106), (33, 0.093), (34, -0.079), (35, 0.026), (36, -0.036), (37, 0.005), (38, -0.014), (39, -0.041), (40, 0.021), (41, 0.005), (42, -0.029), (43, -0.058), (44, -0.062), (45, -0.051), (46, -0.047), (47, -0.037), (48, 0.037), (49, -0.072)]
simIndex simValue paperId paperTitle
same-paper 1 0.9368946 47 iccv-2013-Alternating Regression Forests for Object Detection and Pose Estimation
Author: Samuel Schulter, Christian Leistner, Paul Wohlhart, Peter M. Roth, Horst Bischof
Abstract: We present Alternating Regression Forests (ARFs), a novel regression algorithm that learns a Random Forest by optimizing a global loss function over all trees. This interrelates the information of single trees during the training phase and results in more accurate predictions. ARFs can minimize any differentiable regression loss without sacrificing the appealing properties of Random Forests, like low computational complexity during both, training and testing. Inspired by recent developments for classification [19], we derive a new algorithm capable of dealing with different regression loss functions, discuss its properties and investigate the relations to other methods like Boosted Trees. We evaluate ARFs on standard machine learning benchmarks, where we observe better generalization power compared to both standard Random Forests and Boosted Trees. Moreover, we apply the proposed regressor to two computer vision applications: object detection and head pose estimation from depth images. ARFs outperform the Random Forest baselines in both tasks, illustrating the importance of optimizing a common loss function for all trees.
2 0.80677927 404 iccv-2013-Structured Forests for Fast Edge Detection
Author: Piotr Dollár, C. Lawrence Zitnick
Abstract: Edge detection is a critical component of many vision systems, including object detectors and image segmentation algorithms. Patches of edges exhibit well-known forms of local structure, such as straight lines or T-junctions. In this paper we take advantage of the structure present in local image patches to learn both an accurate and computationally efficient edge detector. We formulate the problem of predicting local edge masks in a structured learning framework applied to random decision forests. Our novel approach to learning decision trees robustly maps the structured labels to a discrete space on which standard information gain measures may be evaluated. The result is an approach that obtains realtime performance that is orders of magnitude faster than many competing state-of-the-art approaches, while also achieving state-of-the-art edge detection results on the BSDS500 Segmentation dataset and NYU Depth dataset. Finally, we show the potential of our approach as a general purpose edge detector by showing our learned edge models generalize well across datasets.
3 0.73496038 352 iccv-2013-Revisiting Example Dependent Cost-Sensitive Learning with Decision Trees
Author: Oisin Mac Aodha, Gabriel J. Brostow
Abstract: Typical approaches to classification treat class labels as disjoint. For each training example, it is assumed that there is only one class label that correctly describes it, and that all other labels are equally bad. We know however, that good and bad labels are too simplistic in many scenarios, hurting accuracy. In the realm of example dependent costsensitive learning, each label is instead a vector representing a data point’s affinity for each of the classes. At test time, our goal is not to minimize the misclassification rate, but to maximize that affinity. We propose a novel example dependent cost-sensitive impurity measure for decision trees. Our experiments show that this new impurity measure improves test performance while still retaining the fast test times of standard classification trees. We compare our approach to classification trees and other cost-sensitive methods on three computer vision problems, tracking, descriptor matching, and optical flow, and show improvements in all three domains.
4 0.70445138 336 iccv-2013-Random Forests of Local Experts for Pedestrian Detection
Author: Javier Marín, David Vázquez, Antonio M. López, Jaume Amores, Bastian Leibe
Abstract: Pedestrian detection is one of the most challenging tasks in computer vision, and has received a lot of attention in the last years. Recently, some authors have shown the advantages of using combinations of part/patch-based detectors in order to cope with the large variability of poses and the existence of partial occlusions. In this paper, we propose a pedestrian detection method that efficiently combines multiple local experts by means of a Random Forest ensemble. The proposed method works with rich block-based representations such as HOG and LBP, in such a way that the same features are reused by the multiple local experts, so that no extra computational cost is needed with respect to a holistic method. Furthermore, we demonstrate how to integrate the proposed approach with a cascaded architecture in order to achieve not only high accuracy but also an acceptable efficiency. In particular, the resulting detector operates at five frames per second using a laptop machine. We tested the proposed method with well-known challenging datasets such as Caltech, ETH, Daimler, and INRIA. The method proposed in this work consistently ranks among the top performers in all the datasets, being either the best method or having a small difference with the best one.
5 0.69966614 437 iccv-2013-Unsupervised Random Forest Manifold Alignment for Lipreading
Author: Yuru Pei, Tae-Kyun Kim, Hongbin Zha
Abstract: Lipreading from visual channels remains a challenging topic considering the various speaking characteristics. In this paper, we address an efficient lipreading approach by investigating the unsupervised random forest manifold alignment (RFMA). The density random forest is employed to estimate affinity of patch trajectories in speaking facial videos. We propose novel criteria for node splitting to avoid the rank-deficiency in learning density forests. By virtue of the hierarchical structure of random forests, the trajectory affinities are measured efficiently, which are used to find embeddings of the speaking video clips by a graph-based algorithm. Lipreading is formulated as matching between manifolds of query and reference video clips. We employ the manifold alignment technique for matching, where the L∞norm-based manifold-to-manifold distance is proposed to find the matching pairs. We apply this random forest manifold alignment technique to various video data sets captured by consumer cameras. The experiments demonstrate that lipreading can be performed effectively, and outperform state-of-the-arts.
6 0.65848523 340 iccv-2013-Real-Time Articulated Hand Pose Estimation Using Semi-supervised Transductive Regression Forests
7 0.65758771 448 iccv-2013-Weakly Supervised Learning of Image Partitioning Using Decision Trees with Structured Split Criteria
8 0.58592039 133 iccv-2013-Efficient Hand Pose Estimation from a Single Depth Image
9 0.52313346 211 iccv-2013-Image Segmentation with Cascaded Hierarchical Models and Logistic Disjunctive Normal Networks
10 0.51660907 443 iccv-2013-Video Synopsis by Heterogeneous Multi-source Correlation
11 0.49089447 165 iccv-2013-Find the Best Path: An Efficient and Accurate Classifier for Image Hierarchies
12 0.48565567 241 iccv-2013-Learning Near-Optimal Cost-Sensitive Decision Policy for Object Detection
13 0.47452796 278 iccv-2013-Multi-scale Topological Features for Hand Posture Representation and Analysis
14 0.46603382 132 iccv-2013-Efficient 3D Scene Labeling Using Fields of Trees
15 0.46431774 136 iccv-2013-Efficient Pedestrian Detection by Directly Optimizing the Partial Area under the ROC Curve
16 0.46325687 176 iccv-2013-From Large Scale Image Categorization to Entry-Level Categories
17 0.45919126 178 iccv-2013-From Semi-supervised to Transfer Counting of Crowds
18 0.44538042 24 iccv-2013-A Non-parametric Bayesian Network Prior of Human Pose
19 0.44535127 391 iccv-2013-Sieving Regression Forest Votes for Facial Feature Detection in the Wild
20 0.4397963 130 iccv-2013-Dynamic Structured Model Selection
topicId topicWeight
[(2, 0.09), (7, 0.026), (12, 0.023), (26, 0.089), (31, 0.029), (40, 0.058), (42, 0.1), (57, 0.242), (64, 0.033), (73, 0.032), (89, 0.141)]
simIndex simValue paperId paperTitle
same-paper 1 0.7589314 47 iccv-2013-Alternating Regression Forests for Object Detection and Pose Estimation
Author: Samuel Schulter, Christian Leistner, Paul Wohlhart, Peter M. Roth, Horst Bischof
Abstract: We present Alternating Regression Forests (ARFs), a novel regression algorithm that learns a Random Forest by optimizing a global loss function over all trees. This interrelates the information of single trees during the training phase and results in more accurate predictions. ARFs can minimize any differentiable regression loss without sacrificing the appealing properties of Random Forests, like low computational complexity during both, training and testing. Inspired by recent developments for classification [19], we derive a new algorithm capable of dealing with different regression loss functions, discuss its properties and investigate the relations to other methods like Boosted Trees. We evaluate ARFs on standard machine learning benchmarks, where we observe better generalization power compared to both standard Random Forests and Boosted Trees. Moreover, we apply the proposed regressor to two computer vision applications: object detection and head pose estimation from depth images. ARFs outperform the Random Forest baselines in both tasks, illustrating the importance of optimizing a common loss function for all trees.
Author: Marco San_Biagio, Marco Crocco, Marco Cristani, Samuele Martelli, Vittorio Murino
Abstract: Capturing the essential characteristics of visual objects by considering how their features are inter-related is a recent philosophy of object classification. In this paper, we embed this principle in a novel image descriptor, dubbed Heterogeneous Auto-Similarities of Characteristics (HASC). HASC is applied to heterogeneous dense features maps, encoding linear relations by covariances and nonlinear associations through information-theoretic measures such as mutual information and entropy. In this way, highly complex structural information can be expressed in a compact, scale invariant and robust manner. The effectiveness of HASC is tested on many diverse detection and classification scenarios, considering objects, textures and pedestrians, on widely known benchmarks (Caltech-101, Brodatz, Daimler Multi-Cue). In all the cases, the results obtained with standard classifiers demonstrate the superiority of HASC with respect to the most adopted local feature descriptors nowadays, such as SIFT, HOG, LBP and feature covariances. In addition, HASC sets the state-of-the-art on the Brodatz texture dataset and the Daimler Multi-Cue pedestrian dataset, without exploiting ad-hoc sophisticated classifiers.
3 0.68129563 358 iccv-2013-Robust Non-parametric Data Fitting for Correspondence Modeling
Author: Wen-Yan Lin, Ming-Ming Cheng, Shuai Zheng, Jiangbo Lu, Nigel Crook
Abstract: We propose a generic method for obtaining nonparametric image warps from noisy point correspondences. Our formulation integrates a huber function into a motion coherence framework. This makes our fitting function especially robust to piecewise correspondence noise (where an image section is consistently mismatched). By utilizing over parameterized curves, we can generate realistic nonparametric image warps from very noisy correspondence. We also demonstrate how our algorithm can be used to help stitch images taken from a panning camera by warping the images onto a virtual push-broom camera imaging plane.
4 0.66222119 442 iccv-2013-Video Segmentation by Tracking Many Figure-Ground Segments
Author: Fuxin Li, Taeyoung Kim, Ahmad Humayun, David Tsai, James M. Rehg
Abstract: We propose an unsupervised video segmentation approach by simultaneously tracking multiple holistic figureground segments. Segment tracks are initialized from a pool of segment proposals generated from a figure-ground segmentation algorithm. Then, online non-local appearance models are trained incrementally for each track using a multi-output regularized least squares formulation. By using the same set of training examples for all segment tracks, a computational trick allows us to track hundreds of segment tracks efficiently, as well as perform optimal online updates in closed-form. Besides, a new composite statistical inference approach is proposed for refining the obtained segment tracks, which breaks down the initial segment proposals and recombines for better ones by utilizing highorder statistic estimates from the appearance model and enforcing temporal consistency. For evaluating the algorithm, a dataset, SegTrack v2, is collected with about 1,000 frames with pixel-level annotations. The proposed framework outperforms state-of-the-art approaches in the dataset, show- ing its efficiency and robustness to challenges in different video sequences.
5 0.6607433 445 iccv-2013-Visual Reranking through Weakly Supervised Multi-graph Learning
Author: Cheng Deng, Rongrong Ji, Wei Liu, Dacheng Tao, Xinbo Gao
Abstract: Visual reranking has been widely deployed to refine the quality of conventional content-based image retrieval engines. The current trend lies in employing a crowd of retrieved results stemming from multiple feature modalities to boost the overall performance of visual reranking. However, a major challenge pertaining to current reranking methods is how to take full advantage of the complementary property of distinct feature modalities. Given a query image and one feature modality, a regular visual reranking framework treats the top-ranked images as pseudo positive instances which are inevitably noisy, difficult to reveal this complementary property, and thus lead to inferior ranking performance. This paper proposes a novel image reranking approach by introducing a Co-Regularized Multi-Graph Learning (Co-RMGL) framework, in which the intra-graph and inter-graph constraints are simultaneously imposed to encode affinities in a single graph and consistency across different graphs. Moreover, weakly supervised learning driven by image attributes is performed to denoise the pseudo- labeled instances, thereby highlighting the unique strength of individual feature modality. Meanwhile, such learning can yield a few anchors in graphs that vitally enable the alignment and fusion of multiple graphs. As a result, an edge weight matrix learned from the fused graph automatically gives the ordering to the initially retrieved results. We evaluate our approach on four benchmark image retrieval datasets, demonstrating a significant performance gain over the state-of-the-arts.
6 0.65158415 326 iccv-2013-Predicting Sufficient Annotation Strength for Interactive Foreground Segmentation
7 0.65157616 180 iccv-2013-From Where and How to What We See
8 0.65033829 421 iccv-2013-Total Variation Regularization for Functions with Values in a Manifold
9 0.64987242 384 iccv-2013-Semi-supervised Robust Dictionary Learning via Efficient l-Norms Minimization
10 0.64954603 318 iccv-2013-PixelTrack: A Fast Adaptive Algorithm for Tracking Non-rigid Objects
11 0.64920294 241 iccv-2013-Learning Near-Optimal Cost-Sensitive Decision Policy for Object Detection
12 0.64816016 197 iccv-2013-Hierarchical Joint Max-Margin Learning of Mid and Top Level Representations for Visual Recognition
13 0.64789093 150 iccv-2013-Exemplar Cut
14 0.64742458 404 iccv-2013-Structured Forests for Fast Edge Detection
15 0.64717346 383 iccv-2013-Semi-supervised Learning for Large Scale Image Cosegmentation
16 0.64614725 245 iccv-2013-Learning a Dictionary of Shape Epitomes with Applications to Image Labeling
17 0.64605695 126 iccv-2013-Dynamic Label Propagation for Semi-supervised Multi-class Multi-label Classification
18 0.64589971 21 iccv-2013-A Method of Perceptual-Based Shape Decomposition
19 0.64589715 6 iccv-2013-A Convex Optimization Framework for Active Learning
20 0.64537477 340 iccv-2013-Real-Time Articulated Hand Pose Estimation Using Semi-supervised Transductive Regression Forests