*5.5. Conclusions*

In this section, four challenges were discussed. For reflective objects, 3D-informationbased approaches cannot accurately measure their pose. For textureless objects, only the geometrical features can be extracted for calculating the 6D pose of objects. Faced with foreground occlusion, many methods use the visible part to describe the invisible part of objects and measure their 6D pose. For background clutter, instance segmentation is used to separate the object, and then the 6D pose is measured. For deformable objects, images of the objects are captured in different poses in different views and a dataset is set up. The best matching is found in the dataset when measuring the pose of objects.

#### **6. Conclusions**

This paper divided solutions of 6D pose estimation into three kinds of approaches and made some detailed introductions to their advantages and disadvantages. Then, this paper focused on the challenges in 6D pose estimations, introducing the difficulties of these problems and summarizing some feasible solutions. Finally, some approaches were qualitatively compared with respect to several different indicators. Learning-based approaches achieved the best robustness. However, they are time-consuming and require lots of storage. 2D-information-based approaches are easy to implement and can be applied online. 3D-information-based approaches can achieve higher accuracy than 2D-information-based approaches, but they require more information (depth information) to be collected and dealt with, and they cannot measure the 6D pose of reflective objects. Generally, the methods presented have already satisfied the requirements of industrial application for the 6D pose detection of general objects. However, these methods cannot maintain their excellent performance under some challenging conditions. In future research, learning-based approaches should be further developed. On one hand, their robust performance should be retained. On the other hand, the offline training time should be decreased. These approaches can also be combined with 2D-information-based and 3D-information-based approaches to obtain more accurate results.

**Author Contributions:** Conceptualization, Z.H. and X.Z.; methodology, Z.H.; software, W.F.; validation, X.Z. and Y.L.; writing—original draft preparation, W.F.; writing—review and editing, Z.H. and X.Z. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research was supported by the National Natural Science Foundation of China (51775498, 51775497) and Zhejiang Province Public Welfare Technology Application Research Project (LGG19E050019).

**Data Availability Statement:** The data presented in this study is contained in the article itself.

**Conflicts of Interest:** The authors declare no conflict of interest.
