Multiview layered depth image

Layered Depth Images (LDI) compactly represent multiview images and videos and have widespread usage in image-based rendering applications. In its typical use case scenario of representing a scanned environment, it has proven to be a less costly alternative than separate viewpoint encoding. However, higher quality laser scanner hardware and different user interaction paradigms have emerged, creating scenarios where traditional LDIs have considerably lower efficacy. Wide-baseline setups create surfaces aligned to the viewing rays producing a greater amount of sparsely populated layers. Free viewpoint visualization suffers from the variant quantization of depths on the LDI algorithm, reducing resolution of the dataset in uneven directions. This paper presents an alternative representation to the LDI, in which each layer of data is positioned in different viewpoints that coincide with the original scanning viewpoints. A redundancy removal algorithm based on world-space distances as opposed to to image-space is discussed, ensuring points are evenly distributed and are not viewpoint dependent. We compared our proposed representation with traditional LDIs and viewpoint dependent encoding. Results showed the multiview LDI (MVLDI) creates a smaller number of layers and removes higher amounts of redundancy than traditional LDIs, ensuring no relevant portion of data is discarded in wider baseline setups.