Disentangling Deep Network for Reconstructing 3D Object Shapes from Single 2D Images

Recovering 3D shapes of deformable objects from single 2D images is an extremely challenging and ill-posed problem. Most existing approaches are based on structure-from-motion or graph inference, where a 3D shape is solved by fitting 2D keypoints/mask instead of directly using the vital cue in the original 2D image. These methods usually require multiple views of an object instance and rely on accurate labeling, detection, and matching of 2D keypoints/mask across multiple images. To overcome these limitations, we make effort to reconstruct 3D deformable object shapes directly from the given unconstrained 2D images. In training, instead of using multiple images per object instance, our approach relaxes the constraint to use images from the same object category (with one 2D image per object instance). The key is to disentangle the category-specific representation of the 3D shape identity and the instance-specific representation of the 3D shape displacement from the 2D training images. In testing, the 3D shape of an object can be reconstructed from the given image by deforming the 3D shape identity according to the 3D shape displacement. To achieve this goal, we propose a novel convolutional encoder-decoder network—the Disentangling Deep Network (DisDN). To demonstrate the effectiveness of the proposed approach, we implement comprehensive experiments on the challenging PASCAL VOC benchmark and use different 3D shape ground-truth in training and testing to avoiding overfitting. The obtained experimental results show that DisDN outperforms other state-of-the-art and baseline methods.