Pose estimation for category specific multiview object localization

By Mustafa Özuysal, Vincent Lepetit, Pascal Fua. 2009

283 citations recorded by the source corpus.

We propose an approach to overcome the two main challenges of 3D multiview object detection and localization: The variation of object features due to changes in the viewpoint and the variation in the size and aspect ratio of the object. Our approach proceeds in three steps. Given an initial bounding box of fixed size, we first refine its aspect ratio and size. We can then predict the viewing angle, under the hypothesis that the bounding box actually contains an object instance. Finally, a classifier tuned to this particular viewpoint checks the existence of an instance. As a result, we can find the object instances and estimate their poses, without having to search over all window sizes and potential orientations. We train and evaluate our method on a new object database specifically tailored for this task, containing real-world objects imaged over a wide range of smoothly varying viewpoints and significant lighting changes. We show that the successive estimations of the bounding box and the viewpoint lead to better localization results.

Subfields: Aerospace Engineering, Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Aerospace Engineering

In Computer Vision and Pattern Recognition

Citations in the corpus

Strike (With) a Pose: Neural Networks Are Easily Fooled by Strange Poses of Familiar Objects (2018, methodology): “Previous research has successfully used 3D graphics as a diagnostic tool for computer vision systems [7, 31, 47, 32, 50].”
Pix3D: Dataset and Methods for Single-Image 3D Shape Modeling (2018, methodology): “Researchers have also proposed numerous methods on 3D pose estimation [13, 42, 55, 58].”
Learning 3D Object Categories by Looking Around Them (2017, background): “The vast majority of methods for learning the viewpoint of object categories use manual supervision [32, 25, 11, 26, 43, 23, 39] or synthetic [36] data.”
Generic 3D Representation via Pose Estimation and Matching (2016, methodology): “Abstraction: Cube Object: we performed a similar abstraction test between a set of 88 convex cubes and the images of EPFL Multi-View Car dataset [43], which includes a dense sampling of various viewpoints of cars in an exhibition.”
Multi-view object class detection with a 3D geometric model (2010, background): “However, these advantages often come with an increased training complexity such as manual perlabel annotations.”
Ieee Transactions on Pattern Analysis and Machine Intelligence 1 Gradient Response Maps for Real-time Detection of Texture-less Objects (background): “For example, in [5], [21], [22], one or several classifiers are trained to detect faces or cars under various views.”
Detailed 3D Representations for Object Recognition and Modeling (2013, methodology): “While viewpoint-invariant detection has been a hot topic for some time now [49], [55], [62], [42], [3], [64], [51], [20], [43], [19], [57], [45], most approaches are made up of several flat viewpoint-dependent representations connected together in one way or the other.”
Discriminative Mixture-of-Templates for Viewpoint Classification (2010, background): “In the case of [30], rotation is limited to in-plane rotation on the ground.”
Are we ready for autonomous driving? The KITTI vision benchmark suite (2012, unknown): “COIL-100 [33] 100 72 X 72 bins EPFL Multi-View Car [34] 20 90 X 90 bins Caltech 3D Objects [31] 100 144 X 144 bins Proposed Dataset 2 80,000 X X continuous”
Cascaded pose regression (2010, methodology): “[25] used pose estimation prior to detection, conditioning their classifier on the pose estimate.”
Teaching 3D geometry to deformable part models (2012, background): “On the technical side, we observe that multi-view recognition is often phrased as a two step procedure, where object localization and viewpoint estimation are performed in succession [24] [20] [13] [36].”
Learning Category-Specific Deformable 3D Models for Object Reconstruction (2017, background): “Index Terms—Object reconstruction, 3D shape modeling, viewpoint estimation, scene understanding Ç”
Towards Viewpoint Invariant 3D Human Pose Estimation (2016, unknown): “[50] trained a classifier for each viewpoint.”
ObjectNet3D: A Large Scale Database for 3D Object Recognition (2016, methodology): “The EPFL Car dataset [23] consists of 2,299 images of 20 car instances at multiple azimuth angles, with almost identical elevation.”
Beyond PASCAL: A benchmark for 3D object detection in the wild (2014, unknown): “PASCAL3D+ (ours) ETH-80 [13] [26] 3DObject [22] EPFL Car [20] [27] KITTI [8] NYU Depth [24] NYC3DCars [19] IKEA [15] Comparison of our PASCAL3D+ dataset with some of the other 3D datasets.”
Deep MANTA: A Coarse-to-Fine Many-Task Network for Joint 2D and 3D Vehicle Analysis from Monocular Image (2017, background): “They generally divide the viewing sphere in several bins to learn multi-class models where each bin corresponds to a class [27, 41, 47, 22, 29].”
Back to the Future: Learning Shape Models from 3D CAD Data (2010, background): “The first major line of research in 3D object recognition starts from a collection of natural training images depicting the object class of interest from varying viewpoints [2, 11, 21, 26, 27].”
Modeling Occlusion by Discriminative AND-OR Structures (2013, background): “A Street Parking Car Dataset There are several datasets featuring a large amount of car images [18, 26, 23, 4, 5], but they are not suitable to evaluating occlusion handling, as the proportion of (moderately or heavily) occluded cars is marginal.”
Vision Meets Drones: A Challenge (2018, unknown): “9k 640× 480 2007 TUD [15] life 1, 818 1 3, 274 640× 480 2008 EPFL Multi-View Car [16] exhibition 2, 000 1 2, 000 376× 250 2009 Caltech Pedestrian [1] driving 249k 1 347k 640× 480 √ 2012 KITTI Detection [2] driving 15.”