Using Stereo for Object Recognition

Scott Helmer and David Lowe

Abstract There has been signicant progress recently in object recognition research, but many of the current approaches still fail for object classes with few distinctive features, and in settings with signicant clutter and viewpoint variance. One such setting is visual search in mobile robotics, where tasks such as nding a mug or stapler require robust recognition. The focus of this paper is on integrating stereo vision with appearance based recognition to increase accuracy and efciency. We propose a model that utilizes a chamfer-type silhouette classier which is weighted by a prior on scale, which is robust to missing stereo depth information. Our approach is validated on a set of challenging indoor scenes containing mugs and shoes, where we nd that priors remove a signicant number of false positives, improving the average precision by 0.2 on each dataset. We additionally experiment with an additional classifer by Felzenszwalb et al.[1] to demonstrate the approachs robustness.

I. I NTRODUCTION Object classication and recognition has progressed rapidly in recent years due to advances in machine learning, more sophisticated feature extraction techniques, and the ever greater availability of image datasets. Despite the recent success, there is still signicant progress required before we have robots assisting the elderly, cleaning our homes, or fetching household items. A particular challenge for mobile robots in an indoor environment is that most of the objects to be manipulated occupy small portions of cluttered scenes. However, much of the success thus far in object recognition/localization has been achieved with large objects that are often assumed to occupy a signicant portion of the image, such as pedestrians, vehicles, and animals [2]. Many smaller objects found within cluttered indoor scenes are left relatively unaccounted for, such as mugs, staplers, shoes, etc. Due to large variations in appearance, these types of object categories are difcult to recognize with patch based methods, which generally require signicant resolution and distinctive features that are internal to the object and therefore not disturbed by background clutter. There is an increasing body of work suggesting that using scene context can improve the efciency and accuracy of localization. Scene cues such as the gist of a scene were successfully used by Torralba [3] to predict the likely location and scale of objects. It has also been shown that local context, such as surrounding image texture can improve object classication and segmentation [4]. Object cooccurrence and co-location have also been shown by many researchers to be valuable in object detection [5], [6], [7].
Scott Helmer and David Lowe are with the University of British Columbia, Department of Computer Science, Vancouver, British Columbia, Canada, [email protected]

More recently, many research results have demonstrated that using 3D scene information such as surface orientation and scale [8], [9], [10], [11], [12], [13] can be useful in object recognition. In the case where the objects in a scene lack sufcient resolution, or are difcult to detect with current recognition methods, scene context can play a large role. The focus in this paper is on fusing 2D image information and depth information from stereo images into one model for localization, particularly in the case of contour-based objects. A primary motivation for our work stems from our recent experiences in designing a vision system for our robot Curious George [14], an entrant in the Semantic Robot Vision Challenge (SRVC). This contest is a visual scavenger hunt, where a robot explores a room and returns a set of images corresponding to a list of objects it was tasked to nd. Despite winning the competition in 2007 and 2008, it was apparent that recognition of generic objects from arbitrary viewpoints is still very much an open challenge. With stereo vision, we can use a prior on object size, which can be as simple as a mean and variance of an objects real world size. We show that this reduces false positives, and can increase computational efciency, which is particularily important with a mobile robot. In conjunction with the prior, we experiment with methods to utilize surface variation with a contour-based classier. The primary contribution of this paper is a model formulation that is robust to missing information from stereo. We validate our approach on a challenging set of scenes containing shoes and mugs. A. Related Work There are a variety of recent approaches that make use of an objects real world scale as a prior for the scale in the image. One of the pioneering works in this regard, and the most similar to our formulation, is that of Hoiem et al.[8]. Using only an image, the approach jointly infers 3D object locations and scene information, such as 3D surface orientation, ground plane, and horizon. Assumptions using the estimated horizon and a prior on an objects real scale are utilized to provide a prior on the expected scale in the image. This prior modulates the response from an appearance-based object detector, showing marked improvement in pedestrian and vehicle localization. A primary distinction between our work and theirs is that our scale prior uses stereo rather than their horizon assumptions, which are not applicable indoors. Another approach that is similar in spirit to our own is the pedestrian detection work of Gavrila and Munder [15]. Here, they utilize sparse stereo and ground plane constraints to determine regions of interest, where they will then run a detector for a pedestrian at the appropriate scale. Similar

Fig. 1.

A stereo camera and monocular camera produce an edge image and a depth map. Missing depth information is shown in white. A contour-based classier evaluates bounding boxes in the edge image. The scale prior weights these scores depending on whether the depth and scale of the bounding box agree, thereby reducing the score of false positives to a much greater degree than true positives.

to our own work, they also make use of a chamfer distance metric. One of the primary distinctions between our work and theirs is that our approach does not require ground plane constraints, and is possibly more robust to missing stereo information. Moreover, our approach follows a clean probablistic framework, where as their approach involves numerous parameters and thresholds that requires extensive training to set. Other recent approaches using object scale in recognition are those of Gould et al.[10], and Quigley et al.[13]. They utilize a mobile robot to acquire high-accuracy depth maps using a laser scanner, which requires 2-3 seconds per scan. From this data, their object detector utilizes both surface variation, 3D shape, and appearance to nd objects such as mugs, cups, staplers, etc. Their system achieves impressive results, but their use of a laser scanner to acquire data is unrealistic for many applications. Our approach utilizes stereo which is faster, cheaper, and less invasive but also requires additional robustness to uncertainty. In regards to object detection, most successful approaches have utilized patch based techniques that decompose recognition into recognizing parts of the object. These range from bag-of-word approaches that discard 2D spatial relationships entirely, to approaches that model the spatial relation between features. However, for classes that are visually dened by their 3D shape, the signal from the identifying contour on a patch is sometimes overwhelmed by the noise of foreground texture and background clutter. There are numerous recent localization approaches, however, that make use of contours and chamfer matching, including [16], [17], [18], [19], [20]. We also experiment with extensions to chamfer matching that utilize depth information. II. M ETHOD The task we are concerned with is localizing an object, obj, in an intensity image Im , while also leveraging a noisy depth image Iz . To achieve this we adopt and adapt a multiscale sliding window classier, a widely used approach to localization. Here, a subset of N windows, {i }i=1..N , are

evaluated to determine a score as to whether they contain the object of interest. We achieve this by using a probability function p(o, |Iz , Im ), where o = {obj, background}. Finally, the scores from the sub windows are combined using non-maximum suppression to determine likely detections. There are two directions from which we can improve object localization: reducing false positives, and increasing the scores for true positives. Given that depth images derived from stereo data are noisy and that the objects are small relative to the depths involved, we cannot place much hope in classication based solely on the depth data. Instead, we focus our efforts on utilizing the depth data to reduce false positives, and to help make the most use of the appearance information. To achieve this we separate information about the appearance, the scale of a scene element (obj or background), and the surface variation. We rst assume that the appearance, Im , and depth information, Iz , are conditionally independent given o. This is not technically true since depth using stereo is derived from appearance information, the effect of this dependence is minimal for recognition. Next, the bounding box implies a centre, scale s , and aspect ratio (which we assume is xed), and we denote Iz () as the depth values within . If we assume that a scene elements presence and appearance are independent of where it is in the scene, then the nature of the elements surface is independent of where it is in the scene. So, if we have a scalar function f (Iz ) that measures where the surface of the element is in the scene, then we can move that surface to the model coordinate system, Iv () = Iz ()f (Iz ). We discuss f (Iz ) further in Section II-A. As stated, now {f (Iz ), Iv (), Im } are statistically independent. So, at a particular bounding box, {f (Iz ), Iv ()} become our features that describe the depths in that bounding box. We dene our score as the probability,

zw =

zc sw zc b bs = d= s d sw

(4)

sw zc
Fig. 2. Scene Geometry

zf zw

Now, in practice we do not know the true values of zw , sw , or d, so we use instead the random variables z, s, d. For a particular point x, dx = d + x , where x is noise due to discretization and the stereo algorithm. We can use the relationships in equation 4 to show, 1 N zx =
x

(6) p(f (Iz ), Iv (), Im |o, )p(o, ) (1) p(Iz , Im ) We assume the errors, {x }x , are zero mean and indep(o|Im , Iv (), )p(f (Iz )|o, )p(, Iv (), Im ) pendent, and by the central limit theorem the second term = p(Iz , Im ) in denominator becomes 0, allowing the approximation in (2) equation 6. Using this, we dene f (Iz ) as the average depth p(o|Im , Iv (), )p(f (Iz )|o, ) (3) in an area within our bounding box, so f (Iz ) zc s/s . The average depth is taken over a bounding box that is times the where Equation 2 follows 1 by independence and an appli- size of the original bounding box , but centered at the same cation of Bayes rule. The nal equation follows because we point in the image, as in Figure 2, a technique also adopted assume uniformity in terms not involving o. by [13]. The reason for this is that depth values around the This formulation is similar in spirit to that of Hoiem edges of will be more likely to fall on the background. In et al.[8], and likewise is general and not dependent upon our experiments we use = 0.8. Formally, if we denote our choice of classier and priors. The rst term is the as this inner bounding box of , and missing depth values object classier which we describe in Section II-B. However, as , then note that the classier can depend on both surface data in Iv () and the intensity image Im , which allows for a more f (Iz ) = mean({Iz (x)|x , x = }) (7) powerful classier based upon both shape and texture. The second term is related to our prior on the scale of a scene The prior for the objects scale can take any form. In element With both these terms, we are primarily concerned this paper we assume that the scale of the object class with detecting the object of interest, obj, so we only search is Gaussian, with parameters {s , s }. This implies that through our p(o, |Iz , Im ), for when o = obj. Its necessary f (Iz ), ie. p(f (Iz )|o, ), is also a Gaussian with parameters to include o in the formulation so that we can utilize the {zc s /s , zc s /s }. It should be noted that if we simply object classier p(o|Im , Iv (), ). tried to use the depth value at the centre of the bounding box, versus an average, the classier barely outperformed the A. Object Scale from Depth base classier, in part because stereo often gives no reliable The scale prior of the object is captured in the distribution values at the centre of a textureless object. p(f (Iz )|o, ), which should also capture the uncertainty Up until this point we have treated the object as plain depth measurements. The scale of an object class could nar. This assumption can be relaxed if the object scales be arbitrarily complex, as could the mechanism that is isometrically and we know the aspect ratio of the objects responsible for errors in depth measurements. In this section dimensions. Here, the distance between the objects centre we rst formulate the prior generically, and then describe the plane, zw , and the frontal plane, zf , will be proportional to approximations made in our approach. the scale for the object, i.e., zw zf = sw . This amounts The geometry of a scene is illustrated in Figure 2. For- to little more than adding s to the mean for f (Iz ). In mally, zw is the distance of the objects centre to the cameras practice, this can be ignored for smaller objects. origin, s is the scale (i.e., height or width) in the image plane, and sw is the scale of the object in the fronto-parallel B. Object Classier p(o, |Iz , Im ) = plane at zw . For the moment we will restrict the object to being represented as a plane parallel to the camera. We denote the focal length of the camera, zc , and the baseline distance between stereo cameras as b. The disparity dx on the image plane for any point x is d (since the object is a plane), and the number of points in the bounding box is N . Using perspective projection and epipolar geometry, the following relationships hold In this section we describe our object classier, which provides p(o|Im , Iv (), ) for equation 3. The object classier is based upon the insight that for many manufactured objects, there are often no distinctive textural features. As a result, we instead base our object classier upon the silhouette of the object since this captures succinctly the shape of the object. The measure that we utilize to determine how close the image Im () is to the object class is based upon an altered

zc bs 1 s b + s N zc s s

(5)

For a bounding box = {s , c }, we scale the edgels in the template by s = s /sm , where sm is the model scale. Using the chamfer score, dsT,E (x), in a logistic function, s our base classier becomes p(o|Im ) = [1 + exp(o 1 dsT,Ie (x )]1 s
Fig. 3. Similarity between the model silhouette T and the edge image E is based upon the sum of the differences (xt )(xe ) , and xt xe 2 for edgels in xt T and their closest matches xe E.

(11)

version of chamfer distance. Although this is not a state ofthe-art classier, it is surprisingly effective for classifying the canonical viewpoints of some contour-based classes. The classier we initially describe here does not utilize the shape information contained in Iv (), we discuss an altered version that does near the end of the section. The chamfer distance was rst introduced by [21] as a means of measuring the distance between two curves. In its most basic form, it is the total distance from all the points in a template point set T to the closest points from point set E. A threshold is used to account for missing edgels. Relative to some translation x applied to T , the thresholded chamfer distance is dened as, dT,E (x) = cham, 1 |T | min(, argmin (xt + x) xe 2 )
xt T xe E

(8) This measure however is inherently biased towards cluttered images [18], where a high density of edgels is likely to have a small chamfer distance despite the fact that the pattern of edgels looks nothing like template T . To overcome this, we take an approach similar to Shotton et al.[16], who added the difference in orientations between matching edgels to the x chamfer distance. Explicitly, if we denote xe t +x to be the xe E closest to xt + x, then we can dene two disjoint x sets, T and T , where T = {xt | xe t +x xt 2 < } and x T = {xt | xe t +x xt 2 }. These sets denote nothing more than splitting T into those edgels with a match less than and those that are too far away. If we let to (x) be the orientation of an edgel modulo , then the orientation penalty is dened as, dT,E orient, (x) = 2 |T | + |T |
x |(xt ) (xe t +x )| (9) xt T

This classier does not make use of the shape information that is available in the depth variances Iv (). We did implement a variation on the score in equation 10, where we added an additional term that penalized deviations in depths for the matching edgels. The intuition here is that since Iv () has mean 0, we expect the variation from zero to be small relative to the objects size. We discuss results with this enhanced version of the classier in the results section as well. There are a number of relevant parameters of the object classier that need to be set or learned. , which modulates the inuence of orientation differences on the distance, was set to 0.25. The parameter was set to 0.15. Both of these parameters are similar to values used by Shotton et al.[16], and were set by cross-validation on an independent dataset. The object silhouette we utilized was acquired taking an image of a prototypical object on an uncluttered background. From this, we extracted the silhouette by using only the contours on the exterior of the object. The parameters o and 1 of the logistic classifer can be learned using maximum likelihood on training data. For the class of mugs, we used Graz 17 [22], and for the shoes we collected a set of training images from the internet. C. Sampling for Detection Our object detector is based upon determining local maxima in the probability function p(o, |Iz , Im ), which amounts to nding the bounding boxes with high scores. In a multiscale sliding window setting, this can be computationally expensive depending upon : 1) the computation required to evaluate a single bounding box, and 2) the sensitivity of the classier to minor changes in scale and location. Computation of a single bounding box is relatively efcient. We use integral images to compute the scale prior and use the distance transform so that chamfer matching is O(k) where k is the number of edgels in our silhouette. In general, the less sensitive to scale and location a classier is, the more sparsely we can sample p(o, |Iz , Im ) and still hope to nd all detections. For location, a large portion of the image can be sampled sparsely since the scale prior and chamfer matching both vary somewhat smoothly. Scale sampling is more tricky, since there can be considerable performance degradation if too few scales are evaluated. For chamfer matching, without scale priors, we found that results were stable when sampling every 1/8 octave in scale space, i.e., rescaling by {2i/8 |i I, 16 < i < 16}. This can be reduced to a sampling rate of 1/4 octave and only sampling at a greater frequency in regions where the chamfer distance is small. However, with a full detector that utilizes the scale prior, the number of samples can be greatly reduced. For example, for shoe detection at a particular location, we

and the total oriented chamfer score is, dT,E (x) = (1 )dcham, + dorient, (10)

where weights the contribution of the orientation difference to the chamfer score. The Figure 3 sheds some light on the oriented chamfer distance. This description thus far has not touched upon the issue of scale. Again, we follow Shotton et al.[16] who scale the template rather than edge image.

could use depth to infer at what scales to sample, allowing a reduction of samples by up to 80 percent. III. E VALUATION To validate our approach we collected a dataset of stereo and still images for a variety of scenes containing mugs and shoes, examples of which can be found in Figure 5. Using these images, we then compare the performance of our chamfer-distance object detector versus the performance when this base detector is augmented with a prior on scale. A. Experimental Setup The intent of our data collection was to produce a set of images that would be challenging for an appearance based classier due to the signicant amount of clutter and textural variation on the object itself. With this in mind, the objects were placed at a variety of depths (1 to 7.5 m), with varying amounts of background clutter. In addition, the shapes and texture of the objects themselves varied. Due to the restricted range of viewpoints covered by our shape model, the objects were placed parallel to the image plane, although the object could be left or right facing. For the mug dataset we collected 20 images from different indoor scenes, with 15 different mugs, with about 3 mugs per scene. For the shoe dataset we also collected 20 images of different scenes, with 8 different shoes, with about 3 shoes per scene. The camera setup consisted of a Canon G7, using 1216x912 images, and a Bumblebee 2 stereo camera, using 1024x768 images, with the Canon camera on top of the Bumblebee as in Figure 1. The stereo algorithm used was Point Greys stereo algorithm provided with the camera, which provides fast, accurate depth maps for textured regions, and annotates ambigous regions as missing information. The motivation for the two camera approach is that the quality of the edge images from the Bumblebee camera were poor in comparison to those of the Canon camera. However, this introduces an additional complication since the 3D point cloud derived from the Bumblebee and its software are in the Bumblebees coordinate system. To overcome this we nd a set of point correspondences between one Bumblebee image and the Canon image using SIFT features and geometric constraints [23]. Using the 3D points from the Bumblebee, we t a projection matrix P , using the Gold standard algorithm [24], that maps all 3D points from the Bumblebee to the Canon. Although this introduces additional errors into the depth image Iz , in practice this produced considerable improvements in both the base classier and the classier that utilized scale as a prior. In the case of shoes for example, the average precision for the base detector on Canon images was 0.49, whereas for the Bumblebee images average precision was 0.35. B. Results and Analysis In order to evaluate our approach, we perform detection over a set of scenes, where any bounding box returned by the system is considered a true positive only if its overlap with the true bounding box is at least 50 percent the area

of the union of the two bounding boxes, which is standard for object detection [2]. The primary metric we utilize for comparison is the average precision (AP). For the shoe dataset, the base classier achieved an AP of 0.51, and an AP of 0.71 with the scale prior. For the mug dataset, the base classier achieved an AP of 0.48 and an AP of 0.72 with the scale prior. Also, as we can see in the recall precision curves in Figure 4, there is a signicant difference in performance in these two classiers. We also experimented with an additional classifer, the deformable parts model (DPM) developed by Felzenswalb et al.[1], which has source code available on the net. In this case, we only trained the classier for mugs since it also coincided with our work for SRVC. Using the output from this classier for p(o Im ), we also found that the results improved with the use of the scale prior on the mugs data set. This can be seen in Figure 4(a), for the DPM models. This demonstrates that the approach is usefull for more than the classier we outlined earlier. The improvement is not as drastic since the DPM is more sophisticated and trained on a large set of training data. There are two types of errors made by the object classier, false negatives and false positives. As can be seen by the shape of the recall-precision curves and from the examples, most of the improvement is a result of fewer false positives. In both object classes, there were a number of instances of false negatives that were due to the failure of the object classier, irregardless of the scale prior. These failures were due partially to failure in edge detection, but also due to the fact that a simple silhouette does not capture object class variation particularly well. We also performed a number of experiments to determine the sensitivity of the approach to other parameters. The rst set of experiments were conducted in regards to the interplay between the parameter settings for the object classier, via 2 o and 1 , and scale prior s . The parameters in the nal results were set by optimizing for the s on a separate 2 dataset, and simply setting the variance, s , to an approximation of the real scale variance of the objects. In subsequent experiments we noticed that the results were fairly robust to alterations in these parameters. For example, if we doubled s , the difference in AP were not signicant. This suggests that the scale prior is primarily removing egregious false positives, not assisting in modeling ne distinctions between detections that have a scale that agrees roughly with depth. We also experimented with utilizing the surface variance, Iv (), i.e., the residuals after the mean in the region was subtracted off, to improve detection. In general, for small objects we do not expect a great deal of surface variation, so discontinuities or cases where the template contour is matching a background segment can be detected using Iv (). In one experiment we embedded Iv () into chamfer matching as mentioned in Section II-B. In another experiment we used a uniform prior on the variation of Iv (), with a range of 0 to object scale. This has the effect of disallowing large discontinuities in depth with the region. In both cases, these approaches did improves results by about 0.025 in AP for

1 0.9 0.8 0.7 0.6 Precision 0.5 0.4 0.3 0.2 0.1 0

Chamfer w/ Scale Prior Chamfer DPM DPM w/ Scale Prior 0 0.1 0.2 0.3 0.4 0.5 Recall 0.6 0.7 0.8 0.9 1

(a) Mugs Fig. 4. Recall Precision Curves

(b) Shoes

both datasets. Again this slight improvement is due to the fact that the scale prior removed a majority of false positives. C. Conclusion and Future Work In this paper we have presented an approach that fuses appearance-based recognition using contours with depth information acquired from stereo. Although previous approaches have made use of depth information for recognition, it has yet to be demonstrated that it is feasible with realistic stereo data in an environment not dependent upon ground plane assumptions, in which noise and missing values can be signicant. As our results indicate, a prior on scale can be utilized to increase the accuracy, efciency, and robustness of object localization on the types of objects expected to be manipulated by mobile robots in indoor environments. In addition, the approach we presented is general in that any object classier can be used. The limitations of a classier based upon an entire silhouette are signicant, including sensitivity to viewpoint, clutter edgels, and intraclass variation in the shape of the object. Future work will focus on utilizing a more sophisticated object classier. In the context of a mobile robot, such as a setting like the SRVC, more sophisticated object class models require more computation, making the use of scale as prior that much more important in focusing attention on relevant regions. Moreover, for shape based objects, the primarily challenges are in disambiguating foreground contours from clutter and modelling variation. Using a scale prior that is independent of the detector can help in reducing false positives, as shown by our results, but provides little help in reducing the noise introduced by clutter edges. Future work will also investigate utilizing depth information to both improve edge detection and assist in reducing the effect of clutter edges on measuring the distance between two curves. R EFERENCES
[1] P. F. Felzenszwalb and D. A. M. andDeva Ramanan, A discriminatively trained, multiscale, deformable part model, in CVPR, 2008. [2] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman, The PASCAL Visual Object Classes Challenge 2008 (VOC2008) Results, http://www.pascalnetwork.org/challenges/VOC/voc2008/workshop/index.html. [3] A. Torralba, Contextual priming for object detection, IJCV, vol. 53(2), pp. 169191, 2003.

[4] J. Shotton, J. Winn, C. Rother, and A. Criminisi, Textonboost: Joint appearance, shape and context modeling for multi-class object recognition, in ECCV, 2006. [5] A. Torralba, K. Murphy, and W. Freeman., Contextual models for object detection using boosted random elds, in NIPS, 2005. [6] D. Parikh, L. Zitnick, and T. Chen., From appearance to context-based recognition: Dense labeling in small images, in CVPR, 2008. [7] S. Kumar and M. Hebert, A hierarchical eld framework for unied context-based classication, in ICCV, 2005. [8] D. Hoiem, A. Efros, and M. Heber, Putting objects in perspective, in CVPR, 2006. [9] B. Leibe, N. Cornelis, K. Cornelis, and L. Gool, Dynamic 3d scene analysis from a moving vehicle, in CVPR, 2007. [10] S. Gould, P. Baumstarck, M. Quigley, , A. Y. Ng, and D. Koller, Integrating visual and range data for robotic object detection, in ECCV Workshop on Multi-camera and Multi-modal Sensor Fusion Algorithms and Applications (M2SFA2), 2008. [11] G. J. Brostow, J. Shotton, J. Fauqueur, , and R. Cipolla, Segmentation and recognition using structure from motion cues, in ECCV, 2008. [12] E. B. Sudderth, A. Torralba, W. T. Freeman, and A. S. Willsky, Depth from familiar objects: A hierarchical model for 3d scenes, in CVPR, 2006. [13] M. Quigley, S. Batra, S. Gould, E. Klingbeil, Q. Le, A. Wellman, and A. Y. Ng, High-accuracy 3d sensing for mobile manipulation: Improving object detection and door opening, in International Conference on Robotics and Automation, 2009. [14] D. Meger, P.-E. Forss n, K. Lai, S. Helmer, S. McCann, T. Southey, e M. A. Baumann, J. J. Little, and D. G. Lowe, Curious george: An attentive semantic robot, Robotics and Autonomous Systems, vol. 56, no. 6, pp. 503511, 2008. [15] S. M. Dariu M. Gavrila, Multi-cue pedestrian detection and tracking from a moving vehicle, IJCV, vol. 73 (1), pp. 4159, 2007. [16] J. Shotton and R. Blake, A.and Cipolla, Multi-scale categorical object recognition using contour fragments, PAMI, 2007. [17] A. Opelt, A. Pinz, and A. Zisserman, A boundary fragment model for object detection, in ECCV, 2006. [18] D. M. Gavrila, A bayesian, exemplar-based approach to hierarchical shape matching, PAMI, vol. 29 (8), pp. 14081421, 2007. [19] M. P. Kumar, P. Torr, and A. Zisserman, Extending pictorial structures for object recognition, in BMVC, 2004. [20] B. Leibe, E. Seeman, and B. Schiele, Pedestrian detection in crowded scenes, in CVPR, 2005. [21] H. Barrow, J. Tenenbaum, R. Bolles, and H. Wolf., Parametric correspondence and chamfer matching: Two new techniques for image matching, in IJCAI, 1977. [22] A. Opelt, A. Pinz, and A. Zisserman, Incremental learning of object detectors using a visual shape alphabet, in CVPR, 2006. [23] D. G. Lowe, Distinctive image features from scale-invariant keypoints, IJCV, vol. 60(2), pp. 91110, 2004. [24] R. I. Hartley and A. Zisserman, Multiple View Geometry in Computer Vision, 2nd ed. Cambridge University Press, ISBN: 0521540518, 2004.

Some examples of mug detections

Some examples of shoe detections