DeepDeform: Learning Non-Rigid RGB-D Reconstruction With Semi-Supervised Data

Figure 1 : We propose a semi-supervised strategy combining self-supervision with sparse annotations to build a large-scale RGB-D dataset of non-rigidly deforming scenes (400 scenes, 390,000 frames, 5,533 densely aligned frame pairs). With this data, we propose a new method for non-rigid matching, which we integrate into a non-rigid reconstruction approach.