DIENTES CON ESQUIRLAS ANTE MORTEM

It is clear from Figure 5.19 that a greater acceleration is achieved for the connec- tion and triple point steps of the algorithm than for creating viewing edges. This is likely because in the viewing edge step long epipolar line segments that intersect many contour MBRs require more intersection tests than the lines in the connec- tions and triple points steps, which generally give rise to very short line segments only intersecting with a single contour bounding box. This analysis suggests that algorithm performance could be further enhanced if the search for viewing edges was replaced by another lightweight and quick to build structure.

5.9.6

Evaluation using humans

The reconstructions of various humans and their associated timings for multi-core CPU processing demonstrates that high quality 3D reconstructions suitable for applications like telepresence can be generated at real-time frame rates.

5.10

Conclusion

The study proposed a method of parallelisation tailored to execution on a single machine and provided indicative results for example multicore CPU and GPU im- plementations. The theoretical standpoint is that by considering parallelisation without distribution, we were able to tailor the parallelisation of each step. As well as this theoretical contribution, this study has provided some useful practical results. It has been shown that the parallel EPVH implementation running on a multi-core CPU of a commodity computer is capable of producing high quality fully textured models of humans captured by 8 high resolution cameras in ap- proximately 25ms, making it highly appropriate for real-time applications such as telepresence. However, we were not able to achieve comparable speed ups from multicore GPUs. Our findings suggest that given enough CPU cores, increas-

ingly complex models could be reconstructed in real-time without the need for network distributed processing. While our study compares results to those of the distributed approach, it does not attempt to reimplement the distributed approach on a single machine. Furthermore it does not attempt to inform how our approach might work across any CPU/GPU implementation.

The implication of our findings is that network distributed processing of the EPVH algorithm can be replaced with multiple core CPU parallel processing, using a scheme that tailors parallelisation to each stage. This simplifies system design, eliminates unnecessary network latency, and increases utilisation of processing resources. In the context of a 3D telepresence system this is an important finding, as network distributed processing creates the opportunity for jitter in latency of an end-to-end system that could confuse the flow of communication.

Chapter 6

Evaluation

This chapter describes the methods and processes by which the suitability of 3D reconstruction in the context of a 3D telepresence system has been evalutated. During the course of this research a sophisticated platform for 3D reconstruc- tion from video was developed that has numerous practical and analytical uses. The features of the platform and how they have been used to develop understand- ing, validate requirements and reduce time in prototyping algorithms and camera placement are first described. Then, by means of two case studies, it is shown how the platform can be used to investigate the impact of camera placement on visual and spatial quality of 3D reconstruction. The first case study begins by investigat- ing the impact of camera placement on the overall spatial constraint of objects, this is followed by a deeper analysis of the impact individual cameras have on the spatial and visual quality of a human head reconstruction. The second case study summarises a collaborative experiment underpinned by the platform, that showed for the first time that 3D reconstruction is able to convey eye gaze to accuracies sufficient for human social interaction.

The work describing the software utility presented in this chapter has been pub- lished in [35], and the collaborative experiment has been published in [104].

6.1

Objectives and research questions

O3: Develop a platform through which the impact of camera placement on spatial

and visual quality of 3D reconstruction can be studied.

Q3: What are the requirements of a system with which the impact of camera

placement on spatial and visual quality can be studied?

Given that for a particular multi-core processor temporal quality of the reconstruc- tion algorithm will depend upon the complexity of inputs provided to it, the ability of the algorithm to achieve interactive frame rates will be largely determined by the numer and resolution of cameras used. Reducing the number of cameras, or their resolution, to achieve real-time performance is likely to impact upon spatial and visual quality of the reconstruction. The spatial constraint achieved by the vi- sual hull is known to improve as the number of cameras increases and their view- points of the object under reconstruction are diversified. The spatial and visual quality of reconstructions is likely to be improved as camera resolution increases. The real-time constraints of a telepresence system lead to a fixed budget in terms of the number of cameras and their resolution for a particular processor. There- fore, achieving the most faithful reconstructions within these constraints requires an understanding of the impact camera placement and resolution have on spatial and visual quality of the output.

6.2

3DRecon, a utility for investigating quality in 3D

reconstruction

In order to develop a deeper understanding of many aspects of the VBR process, 3DRecon was developed. 3DRecon is an interactive utility application that can be used to further understanding of many aspects of the VBR process, including the impact of camera placement on form and texture genesis. The tool allows in

CHAPTER 6. EVALUATION 155 depth analysis of individual frames and sequences of frames forming a 4D ani- mated sequence. Similar tools already exist for the analysis of pre-reconstructed geometry, but to the best of our knowledge this is the first such tool that includes the reconstruction back-end, enabling interactive experimentation with the cam- eraset in terms of which cameras contribute to generation of form and texture. Furthermore, a built in simulator provides an environment for rapid prototyping of algorithms or camera configurations prior to real-world deployment.

The principle features of 3DRecon are:

• Interactive control over camera selection for form and texture genesis.

• A built in simulator enables rapid prototyping using virtual cameras and

synthetic objects.

• Runs from pre-recorded datasets on a filesystem, or with live cameras.

• Provides an algorithm ”plugin” API, enabling rapid prototyping and evalu-

ation of 3D reconstruction algorithms.

6.2.1

Related work

LucyViewer1is an open source application capable of viewing 4D data in a man-

ner similar to 3DRecon. The primary difference being that LucyViewer takes as input pre-reconstructed models in the form of geometry files and textures. Hence LucyViewer can not be used to determine the effect individual cameras have on the form of the reconstructed object. Control over texture application from particular cameras is however possible. Lacking the entire reconstruction backend, LucyViewer is also not able to provide the simulated setting provided by 3DRecon.