BitGrout: Completing a Physical Depth Sensor with a Value-Exact Multiplexer and a Learned INT8 Fill at Sensor Rate

Stereo cameras distinguish between disparities that satisfy internal matching and consistency criteria and image locations for which no estimate is accepted. Learned completion can increase spatial coverage; however, it may overwrite accepted measurements or insert plausible yet geometrically inaccurate values into unresolved regions. This study investigates a constrained alternative in which accepted camera disparities are preserved exactly and learned modification is confined to selected invalid regions. BitGrout integrates a deterministic row-fill, a connected-component area criterion, and a compact INT8 stereo network deployed on a Hailo-10H NPU. Exact preservation is verified from recorded device outputs, while geometric accuracy is evaluated separately for camera-firmware replay and software-stereo inputs. For software stereo on 200 KITTI-2015 pairs, completion decreases mean scene disparity error from 1.989 to 1.661 pixels and the D1 error rate from 12.06\% to 10.80\%. Firmware replay yields heterogeneous results: mean error improves on KITTI, deteriorates on ETH3D, and leads to opposing mean and median conclusions on Middlebury. Three live-camera soak tests sustain approximately 30 frames per second, with a mean capture-to-output latency of approximately 49\,ms. Collectively, these findings establish a deployable completion mechanism with an explicit preservation property while delimiting its accuracy and timing claims. They further demonstrate that spatial coverage, measurement preservation, geometric accuracy, and delivery latency require distinct experimental evidence.