RealSense D405 Depth Accuracy Improvement for Object Sizing/Measurement (Python/ROS2)

We're building a ROS2-based robotics platform that uses an Intel RealSense D405 depth camera for close-range perception and object measurement ("sizing" — estimating real-world dimensions of objects from depth data). We need an experienced RealSense/depth-vision engineer to improve the accuracy, density, and repeatability of our depth output so downstream sizing calculations are reliable. Current setup: Python 3, ROS2 (rclpy), pyrealsense2, OpenCV/NumPy, running on Jetson hardware D405 configured with visual_preset "High Density," gain 16, HDR disabled, emitter always on, 1280x720@30 All of librealsense's built-in post-processing filters (spatial, temporal, hole-filling, decimation) are currently disabled in production; a custom Python node does a simple per-pixel temporal accumulation instead A separate tool deprojects two clicked pixels to 3D and reports the Euclidean distance between them (our current "measurement" path) via rs2_deproject_pixel_to_point We've noticed the depth image is quite sparse at close range (a small fraction of pixels return valid depth per frame), and there are some inconsistencies in how depth units/scale are handled that we'd like audited and fixed Scope of work: Audit our current RealSense camera configuration, filter pipeline, and depth-to-metric conversion for correctness and best practice (including the depth-units inconsistency noted above). Recommend and implement an optimal filter/preset/tuning strategy for the D405 specifically for short-range, high-accuracy sizing use cases (spatial/temporal/hole-filling filters, disparity domain processing, exposure/gain tuning, etc.). Improve depth density and noise characteristics — replace or augment our naive temporal accumulator with a more principled approach (e.g., proper multi-frame fusion, edge-aware filtering, or hybrid depth completion), balancing latency vs. accuracy. Validate accuracy quantitatively: design and run repeatable tests against known-size reference objects/calibration targets at multiple distances, and report measurement error (e.g., ± mm at given ranges). Deliver tuned configuration + updated pipeline code, with before/after accuracy benchmarks. Document the calibration/tuning procedure so our team can maintain it going forward. Requirements: Proven hands-on experience with Intel RealSense SDK (librealsense / pyrealsense2), ideally with D400-series (D405/D415/D435) cameras Solid understanding of stereo depth principles, camera intrinsics/extrinsics, disparity, and depth filtering algorithms Strong Python skills; comfort with ROS2 (rclpy) a strong plus Experience designing accuracy/repeatability validation tests for depth or measurement systems Nice to have: experience with depth completion/sensor fusion (e.g., combining stereo depth with monocular depth models), point cloud processing, or photogrammetry/dimensioning systems Nice to have: embedded/Jetson deployment experience Deliverables: Updated camera config + depth processing code (PR-ready, with clear documentation of changes) Test scripts/data demonstrating measurement accuracy improvement across a specified working distance range Short write-up of findings, chosen approach, and remaining limitations/trade-offs