We demonstrate our algorithm's ability to synthesize extrapolated views by vertically moving the virtual camera off the plane of input cameras.
This works well as long as the reconstructed depth is reasonably accurate (Museum, Statue). For datasets where the input camera moved strictly sideways (e.g., Shop, Guitarist) it is challenging to recover the depth of horizontal lines accurately, and, hence, the results degrade.
The black regions near the top/bottom indicate missing data. Occlusions are not handled correctly during vertical moves (e.g., Conference) since the occlusion behavior of these gradients was not observed in the input images.