Expanding consumer interest in the visual experience, particularly in augmented reality (AR) and virtual reality (VR), has prompted the development of 3D imaging technologies in various forms. Conventional methods for anaglyph image production suffer from static scene assumptions, heavy computational burden and hardware. To alleviate these limitations, we present a novel hybrid architecture that combines the Edge-AI to produce real-time, dynamic anaglyph 3D images employing hardware acceleration of the VLSI for the adaptable depth estimation. Our approach utilizes an embedded CNN model tailored for resource-constrained edge devices. The depth is dynamically predicted for real-time input video frames without relying on multi-view or pre-captured stereo image pairs. Then the generated depth maps dynamically that fast fuse with original RGB frames our proposed four-operand wallace Tree Multiplier on FPGA based VLSI. This multiplier is based on an adaptive compressor logic that can adaptively switch between (7:2, 5:2, 4:2 compressors) based on real-time frame-to-frame motion estimation which can save most of the duplicated computations and logic overhead. Experimental results evaluated on our in-house collected datasets that consist various challenging dynamic scene (i.e., fast moving, illumination variation, and complex background) clearly show the effectiveness of the hybrid system. Preliminary results on the Xilinx UltraScale+ MPSoC show an average frame rate improvement of 271% in comparison to available three-operand VLSI designs. Moreover, extensive objective quality measurements using SSIM, PSNR and LPIPS metrics corroborate significant advancements in perceived depth accuracy and overall image quality. The great reduction of hardware and power consumption, together with dramatic enhancement of real-time performance and image quality, make the proposed method as a smart practical and novel approach for the next-generation AR/VR applications, mobile imaging and autonomous vision modules.
Usha et al. (Tue,) studied this question.