Multi-scale Convolutional Feature Fusion For 6D Pose Estimation
Yan Ren, Jiamin Liu · 2022
In order to obtain accurate pose estimate and satisfy real-time needs, 6D pose estimation of objects is an important task to handle challenging under certain circumstances, such as noisy background, and lighting fluctuations. We propose a 6D pose estimation method with the multi-scale convolutional feature fusion based on the correspondence point method. First, the multi-scale convolutional feature fusion module is added to the encoder-decoder structure corresponding to the extraction of critical points to obtain multi-scale information, and enrich the features extracted from each layer of the network. At the same time, the chain of convolutional layers with residual learning is added in the skip connections of the network to effectively solve the problem of too large a gap between the semantic information of the higher and lower layers. The experimental results demonstrate that the ADD(-S) metric and 2D projection metric on LINEMOD dataset reaches 96.22% and 92.5%, respectively, which are improved compared with the DPOD network based on the correspondence point method. The accuracy is improved by 9.3 and 9.8 percentage points, respectively, for smaller size objects in the dataset. The effectiveness of this method is demonstrated in the 6D pose estimation task based on the corresponding point.