An Experimental Study of Keypoint Descriptor Fusion

Yaling Pan1; Li He2; Yisheng Guan1; Hong Zhang2

10.1109/ROBIO55434.2022.10011825

2022

978-1-6654-8110-6

2022 IEEE International Conference on Robotics and Biomimetics (ROBIO)

699-704

5-9 Dec. 2022

Jinghong, China

Local feature descriptors play a crucial role in computer vision problems, especially robot motion. Existing descriptors are highly accurate, but their performance de-pends on the influence of distracting factors, such as illumi-nation and viewpoint. There is room for further improvement of these descriptors. In this paper, we provide an in-depth analysis of several exciting features of the descriptor fusion model (DFM) we have proposed in our recent work, which uses an autoencoder to combine descriptors and exploit their respective advantages. With this DFM framework, we fur-ther validate that fused descriptors can retain advantageous properties and that our DFM is a generally applicable method with respect to various component descriptors. Specifically, we evaluate multiple combinations of hand-crafted and CNN descriptors concerning their performance on a benchmark dataset with illumination and viewpoint changes to obtain comprehensive experimental results. The results show that the fused descriptors have better matching accuracy than their component descriptors.

Robot motion Computer vision Analytical models Biomimetics Lighting Benchmark testing

其他

IEEE

1.Biomimetic Intelligent Robotics Research Laboratory (BIRL), Guangdong University of Technology, Guangzhou, China
2.Shenzhen Key Laboratory of Robotics and Computer Vision, Southern University of Science and Technology, Shenzhen, China

Yaling Pan,Li He,Yisheng Guan,et al. An Experimental Study of Keypoint Descriptor Fusion[C],2022:699-704.