Citation Context

...object detection for autonomous driving. With the large success of deep learning in the past years, the object detection community shifted from simple appearance scoring on exhaustive sliding windows =-=[1]-=- to more powerful, multi-layer visual representations [2, 3] extracted from a smaller set of object/region proposals [4, 5]. This resulted in over 20% absolute performance gains [6, 7] on the PASCAL V...

Citation Context

...uccess of deep learning in the past years, the object detection community shifted from simple appearance scoring on exhaustive sliding windows [1] to more powerful, multi-layer visual representations =-=[2, 3]-=- extracted from a smaller set of object/region proposals [4, 5]. This resulted in over 20% absolute performance gains [6, 7] on the PASCAL VOC benchmark [8]. The motivation behind these bottom-up grou...

Citation Context

...rful, multi-layer visual representations [2, 3] extracted from a smaller set of object/region proposals [4, 5]. This resulted in over 20% absolute performance gains [6, 7] on the PASCAL VOC benchmark =-=[8]-=-. The motivation behind these bottom-up grouping approaches is to provide a moderate number of region proposals among which at least a few accurately cover the ground-truth objects. These approaches t...

Citation Context

...ve sliding windows [1] to more powerful, multi-layer visual representations [2, 3] extracted from a smaller set of object/region proposals [4, 5]. This resulted in over 20% absolute performance gains =-=[6, 7]-=- on the PASCAL VOC benchmark [8]. The motivation behind these bottom-up grouping approaches is to provide a moderate number of region proposals among which at least a few accurately cover the ground-t...

Citation Context

...resents various biases for grouping, thus yielding multiple diverse solutions [10]. Interestingly, the state-of-the-art R-CNN approach [6] does not work well on the autonomous driving benchmark KITTI =-=[11]-=-, falling significantly behind the current top performers [12, 13]. This is due to the low achievable recall of the underlying box proposals on this benchmark. KITTI images contain many small objects,...

Citation Context

... community shifted from simple appearance scoring on exhaustive sliding windows [1] to more powerful, multi-layer visual representations [2, 3] extracted from a smaller set of object/region proposals =-=[4, 5]-=-. This resulted in over 20% absolute performance gains [6, 7] on the PASCAL VOC benchmark [8]. The motivation behind these bottom-up grouping approaches is to provide a moderate number of region propo...

Citation Context

...uccess of deep learning in the past years, the object detection community shifted from simple appearance scoring on exhaustive sliding windows [1] to more powerful, multi-layer visual representations =-=[2, 3]-=- extracted from a smaller set of object/region proposals [4, 5]. This resulted in over 20% absolute performance gains [6, 7] on the PASCAL VOC benchmark [8]. The motivation behind these bottom-up grou...

Citation Context

...posals are needed to succeed in consequent recognition tasks [21, 22, 7]. [16] introduces learning into proposal generation with parametric energies. Exhaustively sampled bounding boxes are scored in =-=[23]-=- using several “objectness” features. BING [15] proposals also score windows based on an object closure measure as a proxy for “objectness”. Edgeboxes [9] score millions of windows based on contour in...

Citation Context

... using parametric min-cut. The proposed solutions are then scored using simple Gestalt-like features, and typically only 150 top-ranked proposals are needed to succeed in consequent recognition tasks =-=[21, 22, 7]-=-. [16] introduces learning into proposal generation with parametric energies. Exhaustively sampled bounding boxes are scored in [23] using several “objectness” features. BING [15] proposals also score...

Citation Context

...9]. Another successful approach is to frame the problem as energy minimization where a parametrized family of energies represents various biases for grouping, thus yielding multiple diverse solutions =-=[10]-=-. Interestingly, the state-of-the-art R-CNN approach [6] does not work well on the autonomous driving benchmark KITTI [11], falling significantly behind the current top performers [12, 13]. This is du...

Citation Context

...pcdφpcd(x,y) +w > c,fsφfs(x,y) +w > c,htφht(x,y) +w > c,ht−contrφht−contr(x,y) Note that our energy depends on the object class via class-specific weights w>c , which are trained using structured SVM =-=[32]-=- (details in Sec. 3.4). We now explain each potential in more detail. 3 Point Cloud Density: This potential encodes the density of the point cloud within the box φpcd(x,y) = ∑ p∈Ω(y) S(p) |Ω(y)| (1) w...

Citation Context

... holistic model that re-reasons about DPM detections based on priors from cartographic maps. In KITTI, the best performing method so far is the recently proposed 3DVP [12] which uses the ACF detector =-=[30]-=- and learned occlusion patters in order to improve performance of occluded cars. 2 C ar # candidates 10 1 10 2 10 3 10 4 re ca llsatsIo Usth re sh ol ds0. 7 0 0.2 0.4 0.6 0.8 1 BING SS EB MCG MCG-D Ou...

Citation Context

... This is the strategy behind Selective Search [4], which is used in most state-of-the-art detectors these days. Contours in the image have also been exploited in order to locate object proposal boxes =-=[9]-=-. Another successful approach is to frame the problem as energy minimization where a parametrized family of energies represents various biases for grouping, thus yielding multiple diverse solutions [1...

Citation Context

...cclusion levels, demonstrating that our approach produces highly accurate object proposals. In particular, we achieve a 25% higher recall for 2K proposals than the state-of-the-art RGB-D method MCG-D =-=[14]-=-. Combined with CNN scoring, our method outperforms all published results on object detection for Car, Cyclist and Pedestrian on KITTI [11]. Our code and data are online: http://www.cs.toronto.edu/˜ob...

Citation Context

... using parametric min-cut. The proposed solutions are then scored using simple Gestalt-like features, and typically only 150 top-ranked proposals are needed to succeed in consequent recognition tasks =-=[21, 22, 7]-=-. [16] introduces learning into proposal generation with parametric energies. Exhaustively sampled bounding boxes are scored in [23] using several “objectness” features. BING [15] proposals also score...

Citation Context

...n approaches for autonomous driving. In [26], objects are predetected via a poselet-like approach and a deformable wireframe model is then fit using the image information inside the box. Pepik et al. =-=[27]-=- extend the Deformable Part-based Model [1] to 3D by linking parts across different viewpoints and using a 3D-aware loss function. In [28], an ensemble of models derived from visual and geometrical cl...

Citation Context

...Fast R-CNN [34], which share 5 Cars Pedestrians Cyclists Easy Moderate Hard Easy Moderate Hard Easy Moderate Hard LSVM-MDPM-sv [35, 1] 68.02 56.48 44.18 47.74 39.36 35.95 35.04 27.50 26.21 SquaresICF =-=[36]-=- - - - 57.33 44.42 40.08 - - - DPM-C8B1 [37] 74.33 60.99 47.16 38.96 29.03 25.61 43.49 29.04 26.20 MDPM-un-BB [1] 71.19 62.16 48.43 - - - - - - DPM-VOC+VP [27] 74.95 64.71 48.76 59.48 44.86 40.37 42.4...

Citation Context

...works [2, 3], which typically operate on a fixed spatial scope, there has been increased interest in object proposal generation. Existing approaches range from purely RGB [4, 9, 10, 5, 15, 16], RGB-D =-=[17, 14, 18, 19]-=-, to video [20]. In RGB, most approaches combine superpixels into larger regions based on color and texture similarity [4, 5]. These approaches produce around 2,000 proposals per image achieving nearl...

Citation Context

...works [2, 3], which typically operate on a fixed spatial scope, there has been increased interest in object proposal generation. Existing approaches range from purely RGB [4, 9, 10, 5, 15, 16], RGB-D =-=[17, 14, 18, 19]-=-, to video [20]. In RGB, most approaches combine superpixels into larger regions based on color and texture similarity [4, 5]. These approaches produce around 2,000 proposals per image achieving nearl...

Citation Context

...tion. We score bounding box proposals using CNN. Our network is built on Fast R-CNN [34], which share 5 Cars Pedestrians Cyclists Easy Moderate Hard Easy Moderate Hard Easy Moderate Hard LSVM-MDPM-sv =-=[35, 1]-=- 68.02 56.48 44.18 47.74 39.36 35.95 35.04 27.50 26.21 SquaresICF [36] - - - 57.33 44.42 40.08 - - - DPM-C8B1 [37] 74.33 60.99 47.16 38.96 29.03 25.61 43.49 29.04 26.20 MDPM-un-BB [1] 71.19 62.16 48.4...

Citation Context

... They show significant improvements in performance with respect to past work. In [19], RGB-D videos are used to propose boxes around very accurate point clouds. Relevant to our work is Sliding Shapes =-=[25]-=-, which exhaustively evaluates 3D cuboids in RGB-D scenes. This approach, however, utilizes an object scoring function trained on a large number of rendered views of CAD models, and uses complex class...

Citation Context

....08 - - - DPM-C8B1 [37] 74.33 60.99 47.16 38.96 29.03 25.61 43.49 29.04 26.20 MDPM-un-BB [1] 71.19 62.16 48.43 - - - - - - DPM-VOC+VP [27] 74.95 64.71 48.76 59.48 44.86 40.37 42.43 31.08 28.23 OC-DPM =-=[38]-=- 74.94 65.95 53.86 - - - - - - AOG [39] 84.36 71.88 59.27 - - - - - - SubCat [28] 84.14 75.46 59.71 54.67 42.34 37.95 - - - DA-DPM [40] - - - 56.36 45.51 41.08 - - - Fusion-DPM [41] - - - 59.51 46.67 ...

Citation Context

...ly operate on a fixed spatial scope, there has been increased interest in object proposal generation. Existing approaches range from purely RGB [4, 9, 10, 5, 15, 16], RGB-D [17, 14, 18, 19], to video =-=[20]-=-. In RGB, most approaches combine superpixels into larger regions based on color and texture similarity [4, 5]. These approaches produce around 2,000 proposals per image achieving nearly perfect achie...

Citation Context

...l importance in this domain, we place our proposals in 3D and represent them as cuboids. We assume a stereo image pair as input and compute depth via the state-of-the-art approach by Yamaguchi et al. =-=[31]-=-. We use depth to compute a point-cloud x and conduct all our reasoning in this domain. We next describe our notation and present our framework. 3.1 Proposal Generation as Energy Minimization We repre...

Citation Context

...meters are learnt by solving the following optimization problem: min w∈RD 1 2 ||w||2 + C N N∑ i=1 ξi s.t.: wT (φ(x(i),y)− φ(x(i),y(i))) ≥ ∆(y(i),y)− ξi, ∀y \ y(i) We use the parallel cutting plane of =-=[33]-=- to solve this minimization problem. We use Intersectionover-Union (IoU) between the set of GT boxes, y(i), and candidates y as the task loss ∆(y(i),y). We compute IoU in 3D as the volume of intersect...

Citation Context

...ntation Estimation Network We use our object proposal method for the task of object detection and orientation estimation. We score bounding box proposals using CNN. Our network is built on Fast R-CNN =-=[34]-=-, which share 5 Cars Pedestrians Cyclists Easy Moderate Hard Easy Moderate Hard Easy Moderate Hard LSVM-MDPM-sv [35, 1] 68.02 56.48 44.18 47.74 39.36 35.95 35.04 27.50 26.21 SquaresICF [36] - - - 57.3...

Citation Context

...31.08 28.23 OC-DPM [38] 74.94 65.95 53.86 - - - - - - AOG [39] 84.36 71.88 59.27 - - - - - - SubCat [28] 84.14 75.46 59.71 54.67 42.34 37.95 - - - DA-DPM [40] - - - 56.36 45.51 41.08 - - - Fusion-DPM =-=[41]-=- - - - 59.51 46.67 42.05 - - - R-CNN [42] - - - 61.61 50.13 44.79 - - - FilteredICF [43] - - - 61.14 53.98 49.29 - - - pAUCEnsT [44] - - - 65.26 54.49 48.60 51.62 38.03 33.38 MV-RGBD-RF [45] - - - 70....

Citation Context

...works [2, 3], which typically operate on a fixed spatial scope, there has been increased interest in object proposal generation. Existing approaches range from purely RGB [4, 9, 10, 5, 15, 16], RGB-D =-=[17, 14, 18, 19]-=-, to video [20]. In RGB, most approaches combine superpixels into larger regions based on color and texture similarity [4, 5]. These approaches produce around 2,000 proposals per image achieving nearl...

Citation Context

...16 38.96 29.03 25.61 43.49 29.04 26.20 MDPM-un-BB [1] 71.19 62.16 48.43 - - - - - - DPM-VOC+VP [27] 74.95 64.71 48.76 59.48 44.86 40.37 42.43 31.08 28.23 OC-DPM [38] 74.94 65.95 53.86 - - - - - - AOG =-=[39]-=- 84.36 71.88 59.27 - - - - - - SubCat [28] 84.14 75.46 59.71 54.67 42.34 37.95 - - - DA-DPM [40] - - - 56.36 45.51 41.08 - - - Fusion-DPM [41] - - - 59.51 46.67 42.05 - - - R-CNN [42] - - - 61.61 50.1...

Citation Context

... - - SubCat [28] 84.14 75.46 59.71 54.67 42.34 37.95 - - - DA-DPM [40] - - - 56.36 45.51 41.08 - - - Fusion-DPM [41] - - - 59.51 46.67 42.05 - - - R-CNN [42] - - - 61.61 50.13 44.79 - - - FilteredICF =-=[43]-=- - - - 61.14 53.98 49.29 - - - pAUCEnsT [44] - - - 65.26 54.49 48.60 51.62 38.03 33.38 MV-RGBD-RF [45] - - - 70.21 54.56 51.25 54.02 39.72 34.82 3DVP [12] 87.46 75.77 65.38 - - - - - - Regionlets [13]...

Citation Context

...an object closure measure as a proxy for “objectness”. Edgeboxes [9] score millions of windows based on contour information inside and on the boundary of each window. A detailed comparison is done in =-=[24]-=-. Fewer approaches exist that exploit RGB-D. [17, 18] extend CPMC [10] with additional affinities that encourage the proposals to respect occlusion boundaries. [14] extends MCG [5] to 3D by an additio...

Citation Context

... fit using the image information inside the box. Pepik et al. [27] extend the Deformable Part-based Model [1] to 3D by linking parts across different viewpoints and using a 3D-aware loss function. In =-=[28]-=-, an ensemble of models derived from visual and geometrical clusters of object instances is employed. In [13], Selective Search boxes are re-localized using top-down, object level information. [29] pr...

Citation Context

... - - - - - - AOG [39] 84.36 71.88 59.27 - - - - - - SubCat [28] 84.14 75.46 59.71 54.67 42.34 37.95 - - - DA-DPM [40] - - - 56.36 45.51 41.08 - - - Fusion-DPM [41] - - - 59.51 46.67 42.05 - - - R-CNN =-=[42]-=- - - - 61.61 50.13 44.79 - - - FilteredICF [43] - - - 61.14 53.98 49.29 - - - pAUCEnsT [44] - - - 65.26 54.49 48.60 51.62 38.03 33.38 MV-RGBD-RF [45] - - - 70.21 54.56 51.25 54.02 39.72 34.82 3DVP [12...

Citation Context

...ve sliding windows [1] to more powerful, multi-layer visual representations [2, 3] extracted from a smaller set of object/region proposals [4, 5]. This resulted in over 20% absolute performance gains =-=[6, 7]-=- on the PASCAL VOC benchmark [8]. The motivation behind these bottom-up grouping approaches is to provide a moderate number of region proposals among which at least a few accurately cover the ground-t...

Citation Context

...verse solutions [10]. Interestingly, the state-of-the-art R-CNN approach [6] does not work well on the autonomous driving benchmark KITTI [11], falling significantly behind the current top performers =-=[12, 13]-=-. This is due to the low achievable recall of the underlying box proposals on this benchmark. KITTI images contain many small objects, severe occlusion, high saturated areas and shadows. Furthermore, ...

Citation Context

...ans Cyclists Easy Moderate Hard Easy Moderate Hard Easy Moderate Hard LSVM-MDPM-sv [35, 1] 68.02 56.48 44.18 47.74 39.36 35.95 35.04 27.50 26.21 SquaresICF [36] - - - 57.33 44.42 40.08 - - - DPM-C8B1 =-=[37]-=- 74.33 60.99 47.16 38.96 29.03 25.61 43.49 29.04 26.20 MDPM-un-BB [1] 71.19 62.16 48.43 - - - - - - DPM-VOC+VP [27] 74.95 64.71 48.76 59.48 44.86 40.37 42.43 31.08 28.23 OC-DPM [38] 74.94 65.95 53.86 ...

Citation Context

...rk by exploiting the typical sizes of objects in 3D, the ground plane and very efficient depth-informed scoring functions. Related to our work are also detection approaches for autonomous driving. In =-=[26]-=-, objects are predetected via a poselet-like approach and a deformable wireframe model is then fit using the image information inside the box. Pepik et al. [27] extend the Deformable Part-based Model ...

Citation Context

...In [28], an ensemble of models derived from visual and geometrical clusters of object instances is employed. In [13], Selective Search boxes are re-localized using top-down, object level information. =-=[29]-=- proposes a holistic model that re-reasons about DPM detections based on priors from cartographic maps. In KITTI, the best performing method so far is the recently proposed 3DVP [12] which uses the AC...

Citation Context

...27] 74.95 64.71 48.76 59.48 44.86 40.37 42.43 31.08 28.23 OC-DPM [38] 74.94 65.95 53.86 - - - - - - AOG [39] 84.36 71.88 59.27 - - - - - - SubCat [28] 84.14 75.46 59.71 54.67 42.34 37.95 - - - DA-DPM =-=[40]-=- - - - 56.36 45.51 41.08 - - - Fusion-DPM [41] - - - 59.51 46.67 42.05 - - - R-CNN [42] - - - 61.61 50.13 44.79 - - - FilteredICF [43] - - - 61.14 53.98 49.29 - - - pAUCEnsT [44] - - - 65.26 54.49 48....

Citation Context

...34 37.95 - - - DA-DPM [40] - - - 56.36 45.51 41.08 - - - Fusion-DPM [41] - - - 59.51 46.67 42.05 - - - R-CNN [42] - - - 61.61 50.13 44.79 - - - FilteredICF [43] - - - 61.14 53.98 49.29 - - - pAUCEnsT =-=[44]-=- - - - 65.26 54.49 48.60 51.62 38.03 33.38 MV-RGBD-RF [45] - - - 70.21 54.56 51.25 54.02 39.72 34.82 3DVP [12] 87.46 75.77 65.38 - - - - - - Regionlets [13] 84.75 76.45 59.70 73.14 61.15 55.21 70.41 5...

Citation Context

... the wide success of deep networks [2, 3], which typically operate on a fixed spatial scope, there has been increased interest in object proposal generation. Existing approaches range from purely RGB =-=[4, 9, 10, 5, 15, 16]-=-, RGB-D [17, 14, 18, 19], to video [20]. In RGB, most approaches combine superpixels into larger regions based on color and texture similarity [4, 5]. These approaches produce around 2,000 proposals p...

Citation Context

...usion-DPM [41] - - - 59.51 46.67 42.05 - - - R-CNN [42] - - - 61.61 50.13 44.79 - - - FilteredICF [43] - - - 61.14 53.98 49.29 - - - pAUCEnsT [44] - - - 65.26 54.49 48.60 51.62 38.03 33.38 MV-RGBD-RF =-=[45]-=- - - - 70.21 54.56 51.25 54.02 39.72 34.82 3DVP [12] 87.46 75.77 65.38 - - - - - - Regionlets [13] 84.75 76.45 59.70 73.14 61.15 55.21 70.41 58.72 51.83 Ours 93.04 88.64 79.10 81.78 67.47 64.70 78.39 ...