EN FR
Homepage Inria website


Section: New Results

Surface Flow

Recovering dense motion information is a fundamental intermediate step in the image processing chain upon which higher level applications can be built, such as tracking or segmentation. For that purpose, pixel observations in the image provide useful motion cues through temporal variations of the intensity function. We have studied the estimation of dense, instantaneous 3D motion fields over non-rigidly moving surface observed by multi-camera systems. The motivation arises from multi-camera applications that require motion information for arbitrary subjects, in order to perform tasks such as surface tracking or segmentation. To this aim, we have proposed a novel framework that allows to efficiently compute dense 3D displacement fields using low level visual cues and geometric constraints. The main contribution is a unified framework that combines flow constraints for small displacements with temporal feature constraints for large displacements and fuses them over the surface using local rigidity constraints. The resulting linear optimization problem allows for variational solutions and fast implementations. Experiments conducted on synthetic and real data demonstrated the respective interests of flow and feature constraints as well as their efficiency to provide robust surface motion cues when combined.

As an extension of this work, we also studied the situation where a depth camera and one or more color cameras are available, a common situation with recent composite sensors such as the Kinect. In this case, geometric information from depth maps can be combined with intensity variations in color images in order to estimate smooth and dense 3D motion fields. We propose a unified framework for this purpose, that can handle both arbitrary large motions and sub-pixel displacements. The novelty with respect to existing scene flow approaches is that it takes advantage of the geometric information provided by the depth camera to define a surface domain over which photometric constraints can be consistently integrated in 3D. Experiments on real and synthetic data provide both qualitative and quantitative results that demonstrated the interest of the approach[12] .