Impact of Segmentation and Color Spaces in 6D Pose Estimation

Abstract

6D pose estimation is an open challenge due to complex world objects and many possible problems when capturing data from the real world, e.g., occlusions, truncations, and noise in the data. Achieving accurate 6D poses will improve results in other open problems like robot grasping or positioning objects in augmented reality. MaskedFusion is one of the most accurate methods for 6D pose estimation but before estimating the pose, the object needs to be detected and segmented. One of the most important stages in the MaskedFusion 6D pose pipeline is image segmentation because, with good image segmentation, it is possible to discard the background or other non-relevant data that are around the object leaving only the data that are most relevant to the 6D pose estimation. We study the impact of using different image segmentation methods in the MaskedFusion 6D object pose estimation and we also study the impact of the color spaces in the MaskedFusion and DenseFusion methods. The experiments conducted, show how robust MaskedFusion is and that using some filtering operations after the predicted masks improves the accuracy of the method. We also show that, with one of the semantic segmentation methods tested, we achieve on average 97% accuracy on the LineMOD dataset, only 0.2% worst than the baseline that uses the ground truth masks provided by the dataset. With the modifications of the color spaces, we improved MaskedFusion in 1.1% and DenseFusion 0.3% in the LineMOD dataset and reach also 0.3% improvement for the MaskedFusion and 0.4% for the DenseFusion in the YCB-Video dataset.