Improved 6D Pose Estimation Network Based on the ES6D

According to the discontinuity of the threedimensional rotation representation and the singleness of pose estimation method in the ES6D pose estimation network, this paper designs a dual decoder to replace the decoder in the original ES6D. First, this paper make lightweight improvements to ES6D's decoder as an explicit decoder, and use six-dimensional(6D) rotation coefficients to replace quaternions to represent three-dimensional(3D) rotations. Then, this paper designs an implicit decoder with a different working mechanism to assist the training of the supervised explicit encoder. Finally, this paper performs fine-grained learning by freezing the parameters of decoders to enhance pose consistency. Experiments show that the parameters and calculations of the improved decoder in the inference stage are reduced by 33.2% and 33.4% respectively, the inference speed is increased by 20.4%, and the accuracy on the mainstream data sets LineMOD and YCB-Video is increased by 1.0% and 0.6% respectively, and the convergence speed is significantly improved.