Simultaneous Semi-Coupled Dictionary Learning for Matching RGBD Data

Matching with hidden information which is available only during training and not during testing has recently become an important research problem. Matching data from two different modalities, known as cross-modal matching is another challenging problem due to the large variations in the data coming from different modalities. Often, these are treated as two independent problems. But for applications like matching RGBD data, when only one modality is available during testing, it can reduce to either of the two problems. In this work, we propose a framework which can handle both these scenarios seamlessly with applications to matching RGBD data of Lambertian objects. The proposed approach jointly uses the RGB and depth data to learn an illumination invariant canonical version of the objects. Dictionaries are learnt for the RGB, depth and the canonical data, such that the transformed sparse coefficients of the RGB and the depth data is equal to that of the canonical data. Given RGB or depth data, their sparse coefficients corresponding to their canonical version is computed which can be directly used for matching using a Mahalanobis metric. Extensive experiments on three datasets, EURECOM, VAP RGB-D-T and Texas 3D Face Recognition database show the effectiveness of the proposed framework.