Extrusion-Based Primitive Fitting: RGB-D Occluded Object Dataset

This dataset contains 200 single-view RGB-D scenes of real household objects recorded using the onboard RGB-D camera of a TIAGo robot. The robot observes a tabletop workspace from an inclined top–down viewpoint, and each scene includes four real objects selected from a set of 12 instances. The objects are deliberately arranged to create natural partial occlusions between them, providing challenging conditions for single-view 3D perception. The dataset supports research on shape and pose estimation, occlusion handling and primitive-based modeling. For each scene, we provide the corresponding ROS bag file (including point cloud, RGB, depth, and TF data), standalone color and depth images, 2D bounding boxes, object labels, and the estimated superellipsoid parameters obtained using our extrusion-based primitive fitting method. All data are captured in real-world conditions using a physical TIAGo robot and real objects, with natural sensor noise, depth discontinuities, and occlusions. The fitted superellipsoid parameters included in the dataset are algorithmic estimates produced by the extrusion-based reconstruction and optimization pipeline, not ground-truth object geometry. This dataset accompanies the methods presented in:Menendez, E.; Martínez, S.; Díaz-de-María, F.; Balaguer, C. (2024). *Integrating Egocentric and Robotic Vision for Object Identification Using Siamese Networks and Superquadric Estimations in Partial Occlusion Scenarios*. Biomimetics, 9(2), 100. https://doi.org/10.3390/biomimetics9020100