Single-Shot Direct Block Address Encoding for Learning Screen Geometry

Abstract

3D surface reconstruction has many applications in different domains such as projection mapping, virtual reality, robot navigation, human computer interaction and manufacturing inspection, to name a few. Among different methods of 3D reconstruction, structured light is widely used as it is comparatively cheap and accessible and solves the main problem of traditional stereo vision systems which is finding accurate pixel correspondences between two or multiple views. Structured light techniques can be most fundamentally categorized in terms of the number of projected images over time, whether a single image (single-shot) or multiple images (multi-shot). Multi-shot structured light methods take advantage of multiple images that are projected sequentially over time, allowing simple encoding / decoding of projector pixel addresses. In contrast, single-shot structured light is preferred in contexts of dynamically moving cameras, projectors or surfaces, and in scenarios where short projection time is important.

In this thesis, a new framework for designing single-shot structured light images using tag embedding, called Direct Block Address Encoding, is presented which, unlike previous methods, results in efficient encoding, decoding and 3D reconstruction. Also, error detection and correction mechanisms are designed to detect pixel codewords with errors and find their correspondences in the projector image. In addition, the relationship between different design parameters (alphabet size, encoding Scheme, tag size, block size) are derived to cover projectors with different resolutions.

Experimental results demonstrate that the proposed scheme is capable of obtaining projector-camera pixel correspondences at higher speed in comparison with previous tag embedding methods, allowing for learning screen geometry from a single shot with high resolution projectors and dynamic cameras and projectors. The proposed Direct Block Address Encoding scheme offers 2-3 times speed up for 3D reconstruction and 5-6 times speed up for encoding/decoding stages due to not requiring a look-up table and/or an exhaustive search, something not achieved with other methods.