Pose estimation for category specific multiview object localization

Özuysal, Mustafa; Lepetit, Vincent; Fua, Pascal

doi:10.1109/cvprw.2009.5206633

Cited by 80 publications

(136 citation statements)

References 15 publications

Supporting

Mentioning

135

Contrasting

Order By: Relevance

“…For quantitative evaluation, we reported the mean absolute error (MAE) between the estimated and the ground truth poses and the percentage accuracy in terms of AE < 22.5°and AE < 45°(where the absolute error AE = |EstimatedAngle− GroundTruth|) is the same as [6,8]. We represented the input by using the HOG [11] features calculated in the 6 × 6 grids with nine orientation bins for each bounding box, and followed the same setup as the previous approaches [2,6]. Table 1 shows the pose estimation results.…”

Section: Constrained Kernel Regression: Let the Training Set Bementioning

confidence: 99%

“…Experimental results: To evaluate the performance of our CKR model on pose estimation, we performed experiments on three challenging datasets, including the multi-view car dataset [2], the 3D objects dataset [1], the RGB-D dataset [10] and compared them with the state-of-the-art.…”

Section: Constrained Kernel Regression: Let the Training Set Bementioning

confidence: 99%

“…Dense pose estimation: The multi-view car dataset [2] was used for dense pose estimation. For quantitative evaluation, we reported the mean absolute error (MAE) between the estimated and the ground truth poses and the percentage accuracy in terms of AE < 22.5°and AE < 45°(where the absolute error AE = |EstimatedAngle− GroundTruth|) is the same as [6,8].…”

Section: Constrained Kernel Regression: Let the Training Set Bementioning

confidence: 99%

“…a camera looking at the object on a viewing circle. There have been several previous works to solve such a 1D pose estimation problem recently, such as [1][2][3][4][5][6][7][8]. These approaches can be classified into two groups.…”

mentioning

confidence: 99%

“…These approaches can be classified into two groups. One group, including [1][2][3][4][5], solves the problem in a discrete way, i.e. using several discrete (4, 8 or 16) view-based classifiers.…”

mentioning

confidence: 99%

See 4 more Smart Citations

Constrained kernel regression for pose estimation

Zhang

Jiang

2014

Electron. lett.

View full text Add to dashboard Cite

A constrained kernel regression model is proposed to solve the problem of one-dimensional (1D) pose estimation. Unlike the traditional kernel regression model, a circular constraint is applied to the output of the regression function, i.e. using 2D coordinates on a unit circle as output instead of 1D pose angles from 0 to 360°. The experimental results show that with this constraint, the performance of kernel regression on the 1D pose estimation can be improved significantly, and the constrained kernel regression model can run in real-time.Introduction: To automatically recognise the pose/viewpoint of an object is a challenging problem of visual object recognition in the area of computer vision, and pose estimation is also a key issue in many applications, including human-machine interaction, robotics, aerospace etc. Practice people usually recognise objects from several familiar viewpoints, e.g. we recognise people from the front or side view of the face, but not from the back side of the head. For most applications, e.g. a robot finding where the handle of a mug is, or a spacecraft flying around the target spacecraft for rendezvous and docking, onedimensional (1D) pose might be enough. Therefore, in this Letter, we focused on 1D pose estimation, i.e. a camera looking at the object on a viewing circle. There have been several previous works to solve such a 1D pose estimation problem recently, such as [1][2][3][4][5][6][7][8]. These approaches can be classified into two groups. One group, including [1][2][3][4][5], solves the problem in a discrete way, i.e. using several discrete (4, 8 or 16) view-based classifiers. However, the other group, including [6-8], solves pose estimation in a continuous way, practically based on regression frameworks [6,7]. Generally, pose estimation is fundamentally a continuous problem, so we followed up [6-8] estimate continuous 1D poses.Torki and Elgammal [6] learnt a kernel regression (KR) function from an embedded representation of local features in visual images. El-Gaaly et al.[7] used a multi-kernel learning approach to extend the regression framework in [6], called multi-KR in which two kernels were defined for visual inputs and depth inputs separately. Both of them kept the output of their regression function unconstrained, i.e. directly outputting the 1D pose angles from 0 to 360°. The two images with pose angles 1 and 359°s hould be quite similar in visual appearance, but their outputs of the traditional KR function, which are supposed to be 1 and 359°, are extremely different. According to [8], it is clear that the images of one object with 1D pose variation (or captured on a viewing circle) lie on a closed 1D manifold, and such a manifold is homeomorphic to a unit circle. Using the homeomorphic manifold (the unit circle) as a common representation can significantly improve the pose estimation results. However, the model in [8] is generative, and cannot run in realtime, as it needs to minimise the reconstruction error for inference. Hence, we introduce such a circular cons...

show abstract

Section: Constrained Kernel Regression: Let the Training Set Bementioning

confidence: 99%

Section: Constrained Kernel Regression: Let the Training Set Bementioning

confidence: 99%

Section: Constrained Kernel Regression: Let the Training Set Bementioning

confidence: 99%

mentioning

confidence: 99%

mentioning

confidence: 99%

See 3 more Smart Citations

Constrained kernel regression for pose estimation

Zhang

Jiang

2014

Electron. lett.

View full text Add to dashboard Cite

show abstract

Structure from Motion

Argyriou

Rincón

Villarini

et al. 2015

Image, Video &Amp; 3D Data Registration

View full text Add to dashboard Cite

Deep Directional Statistics: Pose Estimation with Uncertainty Quantification

Prokudin

Gehler

Nowozin

2018

Lecture Notes in Computer Science

View full text Add to dashboard Cite

Modern deep learning systems successfully solve many perception tasks such as object pose estimation when the input image is of high quality. However, in challenging imaging conditions such as on low resolution images or when the image is corrupted by imaging artifacts, current systems degrade considerably in accuracy. While a loss in performance is unavoidable, we would like our models to quantify their uncertainty in order to achieve robustness against images of varying quality. Probabilistic deep learning models combine the expressive power of deep learning with uncertainty quantification. In this paper we propose a novel probabilistic deep learning model for the task of angular regression. Our model uses von Mises distributions to predict a distribution over object pose angle. Whereas a single von Mises distribution is making strong assumptions about the shape of the distribution, we extend the basic model to predict a mixture of von Mises distributions. We show how to learn a mixture model using a finite and infinite number of mixture components. Our model allows for likelihood-based training and efficient inference at test time. We demonstrate on a number of challenging pose estimation datasets that our model produces calibrated probability predictions and competitive or superior point estimates compared to the current state-of-the-art.

show abstract

Pose estimation for category specific multiview object localization

Abstract: We propose an approach to overcome the two main challenges of 3D multiview object detection and localization:

Cited by 80 publications

References 15 publications

Constrained kernel regression for pose estimation

Constrained kernel regression for pose estimation

Structure from Motion

Deep Directional Statistics: Pose Estimation with Uncertainty Quantification

Contact Info

Product

Resources

About