Said Homidov

◦ Set the project plan, defined technical requirements, and designed system architecture for the development of camera tracking and depth sensing solutions. ◦ Performed an in-depth review and evaluation of state-of-the-art Visual-Inertial SLAM and Visual-Inertial Odometry systems on mobile devices. ◦ Optimized monocular depth extraction and human segmentation models for mobile inference using TfLite and ONNX, enabling real-time performance on mobile devices. ◦ Designed and developed a multi-threaded, cross-platform Visual-Inertial SLAM system for mobile devices from scratch, featuring instant initialization, 30 fps performance, robust tracking in dynamic scenes, and supporting front and rear-facing cameras. ◦ Integrated Visual-Inertial SLAM with an AR system on iOS and Android, enabling real-time tracking and interaction in augmented reality environments.

◦ Developed depth extraction algorithm for Quad-Pixel sensors in CUDA and OpenMP, achieving 34 ms depth extraction at FHD resolution. ◦ Developed Quad-Pixel disparity un-distortion method, extending the depth range up to 100 cm. ◦ Implemented Quad-Pixel disparity-based alpha matting for foreground segmentation. ◦ Designed a Quad-Pixel sensor-based dual-camera module specifications, achieving a depth range of 5-400 cm. ◦ Implemented efficient Quad-Pixel sensor-based dual-camera depth fusion algorithm, enabling 25 ms depth extraction at FHD resolution. ◦ Developed Python SDK packages for developers. ◦ Designed the system architecture of an instant 3D scanning mobile app on Android using Dual Pixel sensor depth and ARCore. ◦ Developed an efficient dense 3D Reconstruction algorithm for the 3D scanning mobile app - Dex3D, enabling full reconstruction in under 25 seconds on mobile devices. ◦ Implemented meshing and texturing algorithms for the Dex3D mobile app. ◦ Developed depth from small motion algorithm, extending the range of dual-pixel cameras up to 500 cm. ◦ Developed dual-pixel dual-camera depth fusion algorithm.

◦ Proposed Incremental Direct Structure Refinement (IDSR) model to approximate a 3D point re-projection error function using Gaussians. ◦ Designed and implemented the Sparse-IDSR algorithm in COLMAP, optimizing the Bundle Adjustment solver and achieving a 4× runtime reduction. ◦ Developed the Dense-IDSR algorithm for efficient multi-view fusion of depth maps from Offset-Pixel Aperture cameras, reducing RMSE by 3×.