Challenges in Embedded Vision for Augmented Reality - Advances in Embedded Computer Vision - page 175

Fig. 7.16 Processor architecture gains in performance per power consumption
AR, the most performance-consuming algorithms, as we have seen before, are related
to 3D object detection, matching, and tracking, which highly benefits from parallel
processing. Different processor architectures enable different gains in performance
per power consumption (MIPS/mW) as shown in Fig. 7.16 . It is highly desirable to
have a SIMD Vector processor or a dedicated HW engine available for AR algorithm
processing to get to the needed performance/power level and to keep the GPU free
for rendering the virtual overlays (Fig. 7.17 ).
Fig. 7.17 Architecture of the Metaio AR Engine