FMTT : Fused Multi-head Transformer with Tensor-compression for 3D Point Clouds Detection on Edge Devices

The real-time detection of 3D objects represents a grand challenge on edge devices. Existing 3D point clouds models are over-parameterized with heavy computation load. This paper proposes a highly compact model for 3D point clouds detection using tensor-compression. Compared to conventional methods, we propose a fused multi-head transformer tensor-compression (FMTT) to achieve both compact size yet with high accuracy. The FMTT leverages different ranks to extract both high and low-level features and then fuses them together to improve the accuracy. Experiments on the KITTI dataset show that the proposed FMTT can achieve 6.04x smaller than the uncompressed model from 55.09MB to 9.12MB such that the compressed model can be implemented on edge devices. It also achieves 2.62% improved accuracy in easy mode and 0.28% improved accuracy in hard mode.