MFF6D: Multi-Scale Feature Fusion Network for 6D Object Pose Estimation and Refinement

Due to insufficient correlation between features at different scales, the multi-scale 6D object pose estimation remains a challenge. To fully correlate the multi-scale features of different modes, we propose a high-efficiency Multi-scale Feature Fusion Network for 6D object pose estimate. Our method utilizes two independent branches to extract appearance features and geometric features respectively, and a multi-scale architecture to focus on semantically rich regions in different scales. In addition, we also propose adaptive channel attention to refine multi-scale features by constraining the average and maximum pooling layers with a trainable parameter. We evaluate our model on the well-known YCB-Video and LineMOD datasets. Our method achieves 94.3% of the AUC on the YCB-Video dataset and 98.3% of the ADD on the LineMOD dataset.