Chinese Journal of Liquid Crystals and Displays, Volume. 37, Issue 7, 913(2022)
Object 6D pose estimation algorithm based on improved heatmap loss function
Fig. 6. Qualitative results for single object pose estimation（green 3D bounding boxes represent the ground truth poses，and blue 3D bounding boxes represent our predictions）
Fig. 7. Relationship between the keypoint error and the distance between the 3D keypoint and the object center
Fig. 8. Qualitative results for occluded object pose estimation（green 3D bounding boxes represent the ground truth poses，and blue 3D bounding boxes represent our predictions）
Table 1. Single object pose estimation ADD（-S）metric test results
View tableView in ArticleTable 1. Single object pose estimation ADD（-S）metric test results
算法 Ours ［ 7 ］［ 6 ］［ 9 ］［ 16 ］［ 16 -17 ］ape 64.8 41.2 21.6 43.6 - 77.0 bench 99.2 85.7 81.8 99.9 - 97.5 cam 88.3 78.9 36.6 86.9 - 93.5 can 96.8 85.2 68.8 95.5 - 96.5 cat 82.7 73.9 41.8 79.3 - 82.1 driller 98.0 77.0 63.5 96.4 - 95.0 duck 60.1 42.7 27.2 52.6 - 77.7 eggbox* 98.9 78.9 69.6 99.2 - 97.1 glue* 99.1 72.5 80.0 95.7 - 99.4 hole 78.4 63.9 42.6 81.9 - 52.8 iron 96.8 94.4 75.0 98.9 - 98.3 lamp 97.3 98.1 71.1 99.3 - 97.5 phone 93.6 51.0 47.7 92.4 - 87.7 平均值 88.8 72.6 56.0 86.3 62.7 88.6
Table 2. Keypoint positioning error and pose estimation error
View tableView in ArticleTable 2. Keypoint positioning error and pose estimation error
/cm /pixel /cm /（°） ape 9.74 1.47 0.95 2.76 bench 28.69 2.90 0.87 1.84 cam 17.15 2.02 0.84 1.97 can 19.34 2.84 0.80 1.32 cat 15.26 2.11 0.90 2.33 duck 25.94 1.78 1.08 2.76 driller 10.71 4.72 0.87 1.76 eggbox* 17.63 3.93 1.35 1.69 glue* 16.48 2.23 1.25 2.58 hole 14.82 2.29 1.00 2.37 iron 30.31 5.17 1.13 2.37 lamp 28.55 5.12 1.24 2.19 phone 20.83 2.74 0.92 2.18 平均值 / 3.02 1.02 2.16
Table 3. Occluded object pose estimation ADD（-S）metric test results
View tableView in ArticleTable 3. Occluded object pose estimation ADD（-S）metric test results
算法 Ours ［ 18 ］［ 6 ］［ 9 ］［ 16 ］［ 16 -17 ］ape 23.6 17.6 2.5 15.8 9.9 59.2 can 55.6 53.9 17.5 63.3 45.5 63.5 cat 16.1 3.3 0.7 16.7 0.8 26.2 driller 59.5 19.2 1.1 65.7 41.6 55.6 duck 33.7 62.4 7.7 25.2 19.5 52.4 glue* 48.3 39.6 10.1 49.6 46.2 71.7 hole 47.3 21.3 5.5 39.7 27.0 52.5 平均值 40.6 31.0 6.4 39.4 27.2 54.4
Table 4. Comparison results of comparative experiment ADD（-S）metric test
View tableView in ArticleTable 4. Comparison results of comparative experiment ADD（-S）metric test
损失函数 HWing Loss MSE Loss HWing Loss MSE Loss 训练轮次 30 30 10 10 ape 64.8 47.3 56.2 41.2 bench 99.2 92.4 98.4 47.7 cam 88.3 85.2 84.5 64.4 can 96.8 93.5 96.9 83.9 cat 82.7 69.4 79.5 62.1 driller 98.0 91.9 95.1 33.6 duck 60.1 44.5 53.8 88.6 eggbox* 98.9 98.4 97.7 66.9 glue* 99.1 92.8 91.4 86.9 hole 78.4 73.6 71.9 57.2 iron 96.8 95.8 96.0 93.1 lamp 97.3 90.7 92.6 86.9 phone 93.6 87.1 92.5 80.4 平均值 88.8 81.7 85.1 68.7
Table 5. Influence of different loss function on the error of predicting HeatMap
View tableView in ArticleTable 5. Influence of different loss function on the error of predicting HeatMap
损失函数 HWing Loss MSE Loss 训练轮次 30 10 30 10 前景像素误差 0.139 9 0.151 5 0.161 6 0.207 7 全局像素误差 0.007 4 0.009 1 0.001 1 0.001 9
Get Citation
Copy Citation Text
Lin LIN, Yan-jie WANG, Hai-chao SUN. Object 6D pose estimation algorithm based on improved heatmap loss function[J]. Chinese Journal of Liquid Crystals and Displays, 2022, 37(7): 913
Category:
Received: Dec. 3, 2021
Accepted: --
Published Online: Jul. 7, 2022
The Author Email: Yan-jie WANG (wangyj@ciomp.ac.cn)
