HandO: a hybrid 3D hand–object reconstruction model for unknown objects

In various multimedia applications, it is of great significance to reconstruct 3D meshes of hands and objects from single RGB images. Mesh-based methods mainly resort to mesh displacements by estimating relative positions between hands and objects, while the distance may be inaccurate. Methods based on signed distance function (SDF) learn relative positions by concurrently sampling hand meshes and object meshes; unfortunately, these methods have very limited capability of reconstructing smooth surfaces with rich details. For example, SDF-based methods are inclined to lose the typologies. To the best of our knowledge, only limited works can simultaneously reconstruct the hands and objects with smooth surfaces and accurate relative positions. To this end, we present a novel hybrid model—hand–object Model (HandO) enabling the hand–object 3D reconstruction with smooth surfaces and accurate positions. Critically, our model for the first time makes the hybrid 3D representation for this task by bringing meshes, SDFs, and parametric models together. A feature extractor is employed to extract the image features, and SDF sample points are projected onto these features to extract the local features of each sampled point. Essentially, our model can be naturally extended to reconstruct a whole body holding an object via the new hybrid representation. Additionally, to overcome the lack of training data, a synthetic body-holding dataset is contributed to the community, thus facilitating the research of reconstructing the hand and object. It contains 31763 images of over 50 object categories. Extensive experiments demonstrate that our model can achieve better performance over the competitors on benchmark datasets.