Automatically generated depth maps from video are usually not aligned with the objects in the original image and produced at lower resolutions. We propose to apply a joint-bilateral filter to smoothen the depth map within the objects and upsample it to the original image resolution, while keeping the object edges in the depth map aligned with the original image. We performed algorithmic and ASIC specific optimizations to achieve a real-time implementation in dedicated hardware, while preserving high quality results. Upsampling depth maps from 240 ÃÂ 135 @ 60 Hz to 1920 ÃÂ 1080 @ 60 Hz, requires 0.35 mm2, at 200 MHz, for a dedicated hardware implementation in CMOS 45 nm technology.
展开▼