6DoF pose 估计evaluation metrics 汇总，持续更新~~

淩亂°似流年 2022-10-11 13:42 239阅读 0赞

**转载请注明作者和出处： [http://blog.csdn.net/john\_bh/][http_blog.csdn.net_john_bh]**

### 文章目录 ###

*   *  1. 2D-projection error
     *  2. ADD
     *  3.ADD-S
     *  4. Visible Surface Discrepancy (VSD)

## 1. 2D-projection error ##

> 2D-projection error measures the average distance between the 2D projections in the image space of the 3D model points, transformed using the ground-truth pose and the predicted pose. The pose estimate is considered to be correct if it is within a selected threshold. 2D-Proj denotes the percentage of correctly estimated poses using a 2D Projection Error threshold set to 5 pixels. For symmetric objects, the 2D projection error is computed against all possible ground truth poses, and the lowest value is used.

2D-projection error 是测量 3D 模型点图像空间中 2D 投影之间的平均距离，使用 ground truth 姿势和预测姿势进行转换。 如果姿态估计在选定的阈值内，则认为姿态估计是正确的。 `2D-Proj 表示使用 2D 投影误差阈值设置为 5 像素的正确估计姿势的百分比`。 对于对称对象，2D 投影误差是针对所有可能的 ground truth姿态计算的，并使用最小值。

P r o j . 2 D = 1 m ∑ x ∈ M ∣ ∣ K ( R x + t ) − K ( R ^ x + t ^ ) ∣ ∣ Proj.2D = \\frac\{1\}\{m\}\\sum\_\{x\\in M\}||K(Rx+t)-K(\\hat Rx+\\hat t)|| Proj.2D=m1x∈M∑∣∣K(Rx\+t)−K(R^x\+t^)∣∣

## 2. ADD ##

paper：[Model based training, detection and pose estimation of texture-less 3D objects in heavily cluttered scenes][Model based training_ detection and pose estimation of texture-less 3D objects in heavily cluttered scenes]

> ADD measures the average distance between the 3D model points transformed using the ground-truth pose and the predicted pose.  
> For ADD metric, a pose is regarded as correct if the average distance of model points is less than 10% of the model diameter.

ADD测量使用 ground truth 姿势和预测姿势转换的 3D 模型点之间的平均距离。对于 ADD 度量，`如果模型点的平均距离小于模型直径的 10%，则认为姿势正确`。

A D D = 1 m ∑ x ∈ M ∣ ∣ ( R x + t ) − ( R ^ x + t ^ ) ∣ ∣ ADD = \\frac\{1\}\{m\}\\sum\_\{x\\in M\}||(Rx+t)-(\\hat Rx+\\hat t)|| ADD=m1x∈M∑∣∣(Rx\+t)−(R^x\+t^)∣∣

## 3.ADD-S ##

RSS 2018：[PoseCNN: A Convolutional Neural Network for 6D Object Pose Estimation in Cluttered Scenes ][PoseCNN_ A Convolutional Neural Network for 6D Object Pose Estimation in Cluttered Scenes]

A D D = 1 m ∑ x 1 ∈ M min ⁡ x 2 ∈ M ∣ ∣ ( R x 1 + t ) − ( R ^ x 2 + t ^ ) ∣ ∣ ADD = \\frac\{1\}\{m\}\\sum\_\{x\_1\\in M\}\\min \\limits\_\{x\_2\\in M\}||(Rx\_1+t)-(\\hat Rx\_2+\\hat t)|| ADD=m1x1∈M∑x2∈Mmin∣∣(Rx1\+t)−(R^x2\+t^)∣∣

## 4. Visible Surface Discrepancy (VSD) ##

ECCV 2016：[On evaluation of 6d object pose estimation ][On evaluation of 6d object pose estimation]  
[paper 地址2][paper _2]  
[PPT][]  
[github][]

> Visible Surface Discrepancy (VSD) is an ambiguityinvariant metric that depends on the visible surface, which can account for the symmetry better.

[http_blog.csdn.net_john_bh]: http://blog.csdn.net/john_bh/
[Model based training_ detection and pose estimation of texture-less 3D objects in heavily cluttered scenes]: http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.250.6547&rep=rep1&type=pdf
[PoseCNN_ A Convolutional Neural Network for 6D Object Pose Estimation in Cluttered Scenes]: https://arxiv.org/abs/1711.00199
[On evaluation of 6d object pose estimation]: https://openreview.net/forum?id=H1WS89buWr
[paper _2]: http://www.v3c.cz/download/vysledky-faze2/V009-2016-Hodan.pdf
[PPT]: https://pdfs.semanticscholar.org/fc4b/d3ba3850f4d4f5d9b663824b007ccba755d3.pdf
[github]: https://github.com/thodan/obj_pose_eval