Modified-EfficientPose: Modifying the Prediction Network of the EfficientPose Method to Improve 6D Pose Estimation Performance
Budi Sulistiyo Nugroho, Anny Yuniarti · Nusantara Science and Technology Proceedings · 2025
In the development of various smart technologies, the 6D pose estimation approach has a very important role in identifying the position of objects in 3D space. One of the state-of-the-art methods for estimating 6D poses is EfficientPose. In single-object conditions, this method shows optimal performance. However, in the case of multiple objects, this method is still not optimal. Multiple objects condition provides complicated problems in 6D Pose estimation, making it a major challenge in the current study. This study is intended to improve the performance of the EfficientPose method by modifying the prediction network section. The prediction network has an important role in processing feature representations into estimation results. One of the important layers in this prediction network is the convolution process. We modify this convolution section to produce better prediction performance. The EfficientPose method uses depthwise separable convolution. The use of the convolution approach is generally done to increase efficiency. The use of normal convolution is done to get better performance, although in terms of efficiency, it decreases. In the study, we use normal convolution. We name the modified result of the EfficientPose method as Modified-EfficientPose. The experiment is conducted using a standard dataset, namely LineMod-Occluded. Meanwhile, the performance of 6D pose estimation is measured using the Average Distance Diameter (ADD) metric. The experimental results show that the performance of the Modified-EfficientPose method (79.11%) has better performance than the EfficientPose method (79.04%), with a difference of only 0.07%. This performance improvement, although very small, is an important step in the development of better methods in the future.