HO-3D Hand-Object Interaction 3D Image Dataset

HO stands for Hand and Object Poses, which is a dataset of annotated 3D images of hand-object interactions. The dataset contains 66,034 training images (from 55 sequences) and 11,524 test images (from 13 sequences). The sequences were acquired by multi-camera and single-camera, and contain 10 objects and 10 objects from the YCB dataset. The annotations are automatically generated by an optimization algorithm. The gesture annotations of the test set are retained. The accuracy of the algorithms in the test set can be evaluated by using standard metrics using the CodaLab challenge (see the project page). The object pose annotations of the test and training sets are publicly available.

The dataset can be used to develop algorithms for predicting hand pose when objects are severely occluded based on a single RGB image approach, and can be generalized to hand pose image prediction for objects not included in the dataset.