%0 Generic %A Jalba, Andrei %A Bertens, Tijn %D 2025 %T Annotated SYMSOL II extension dataset from the paper "Symmetry-induced ambiguity in orientation estimation from RGB images" %U %R 10.4121/415c5a6e-1f17-440e-8e1f-753f90732f9b.v1 %K symmetric solids %K symmetry %K ambiguity %K pose estimation %K computer vision %K machine learning %X
This is the annotated SYMSOL II extension from the paper "Symmetry-Induced ambiguity in orientation estimation from RGB images". Contained are 45000 training images and 5000 test images for both marked cylinder and marked tetrahedron objects. Test images are annotated with full symmetry sets. Training images are annotated only with the groundtruth orientation.
Data is separated into two directories: "/train/" and "/test/" holding training data and test data respectively. Both directories are further subdivided into "/cylO/" and "/tetX/" directories for the marked cylinder and marked tetrahedron respectively.
Annotations are stored as lists of numpy arrays in pickle files (.pkl). The annotation for a given image is stored at the corresponding index in the annotation list. In the training data, each image is annotated with a single 3x3 ground truth rotation matrix. In the test data, each image is annotated with an Nx3x3 array of rotation matrices, where N is the number of rotations in the symmetry set. Continuous symmetry sets for the marked cylinder are sampled at a 1-degree angular resolution.