How can you easily create compelling three-dimensional (3D) images from single, conventional two-dimensional (2D) photographs? Shesh et al. address this question and propose an innovative solution in the context of out of bounds imagery.
Out of bounds imagery is a relatively recent technique that exploits two strong cues--occlusions and shadows--to generate depth-rich images from single photographs. This is achieved by adding 3D objects in between different layers of depth of the scene captured in the image, which generates new occlusions, and by adding shadows cast from the 3D objects that increase the sensation of depth in conventional 2D images. This technique, although successful and widely used, still poses several difficulties, such as the addition of 3D objects to the image so that it conforms with the geometry of the scene and the correct casting of shadows for any added 3D objects illuminated by an arbitrary source.
The paper presents a flexible, interactively driven workflow that attempts to solve the aforementioned issues. First, a 3D frame is interactively placed in the image by dragging its four corners, followed by a calibration process that ensures its correct placement in the scene. Second, the foreground objects are extracted in the form of alpha matte, using GrabCut, the well-established segmentation technique based on graph cuts. Finally, shadows for the 3D frame and foreground objects are generated; these provide strong visual cues that anchor the object onto the ground plane and enhance the composition’s overall realism.
Shesh et al. present a variety of impressive results that verify the validity of their approach and provide a comparison with other existing techniques. The interested reader can find in the paper further details, including the proposed closed-form, real-time solution for the camera calibration and the algorithm for casting the shadows of the foreground objects.