MAXSCAN
Realtime Dense & Sparse Reconstruction and visual localization for AR.
Abstract
Designed and implemented a cross-platform mobile mapping system featuring:
- Real-time sparse SLAM and localization on Android and iOS
- Dense volumetric reconstruction using LiDAR on iOS devices
- GPU-accelerated TSDF integration and surface extraction using Swift + Metal
- Modular architecture for integration with AR frameworks and scalable map management
Problem
Mobile XR applications demand high-performance mapping, but:- Sparse and dense reconstruction are often separated, with limited real-time capability on mobile hardware
- Dense reconstruction on iOS requires low-level GPU programming and sensor optimization
- On Device localization and persistent map merging are technically complex and poorly supported in most toolchains
Contribution
- Architected a modular, scalable SLAM system for mobile, supporting multiple reconstruction modes (sparse/dense)
- Implemented GPU-accelerated TSDF pipeline using Metal for dense LiDAR reconstruction
- Developed a localization and map-merging engine using lightweight data structures suitable for mobile memory constraints
- Tuned system to run at real-time frame rates by optimizing memory access patterns and Metal shader performance
- Integrated the system into mobile AR frameworks for live deployment and testing
Result
- Achieved real-time sparse SLAM and localization across Android and iOS devices
- Enabled dense 3D reconstruction using LiDAR at high frame rates on iOS
- Delivered <1s re-localization latency and consistent map merging across sessions
- Deployed internally as a 3D space scanning tool for persistent AR use cases