An Efficient Multi-view 3D Object Recognition Mechanism for Distributed Edge Devices

This paper proposes an efficient multi-view 3D object recognition mechanism for distributed resource-constrained edge devices. Multi-view based vision processing has various applications on resource-constrained edge devices, like collaborative drone swarms, distributed smart cameras etc. However, huge data transmission from the different edge devices leads to this solution suffering from large bandwidth consumption, long latency, and energy consumption. The efficient multi-view 3D object recognition mechanism includes not only a feature map enhanced scheme, which exploits semantic information to bolster the useful attributes and suppress those less important feature maps, but also the feature maps selection schemes to decrease the amount of data transmission. The proposed fuses and transmits the enhanced feature map, achieves the almost same recognition accuracy with 9-12 times data reduction that can save power consumption and transmission bandwidth compared to the traditional multi-view 3D object recognition mechanism.