Pose Estimation by Removing Negative Surface on the Next Viewpoint

We have developed a new pose estimation neural network that focuses on surfaces that become newly visible with the viewpoint movement. Predicting the next viewpoint for pose estimation and moving there improves pose estimation accuracy. However, when there are a few newly visible surfaces after viewpoint movement, the accuracy improvement is limited. To address this, we have developed a novel method making masked images removing negative impact surfaces that was visible before viewpoint movement, allowing the neural network to focus on the newly visible surfaces. As a result, pose estimation accuracy improved. Experimental results showed that using the masked image improved the success rate of pose estimation by 3.4 points compared to the conventional method without the masked image. Furthermore, our method enables the robot to display 89.5% of products, which is an improvement of 5.3 points over the previous method.