FlashSLAM is a cutting-edge SLAM (Simultaneous Localization and Mapping) system designed to enable efficient and robust 3D scene reconstruction in real-time. By leveraging 3D Gaussian Splatting (3DGS) and advanced vision-based camera tracking, FlashSLAM overcomes the limitations of existing methods in sparse view settings and during large camera movements.
Key innovations include:
- Fast and Accurate Camera Tracking: Achieves sub-80 ms pose estimation using a pretrained feature matching model and point cloud registration, offering a 90% reduction in tracking time compared to SplaTAM.
- Robustness to Sensor Noise: Effectively handles depth errors, enabling reliable performance with consumer-grade devices such as smartphones.
- Improved Accuracy in Sparse Settings: Delivers up to a 92% improvement in tracking accuracy under challenging conditions.
This repository includes the codebase, evaluation scripts, and dataset configurations for replicating the results presented in our paper.
- Real-Time Performance: Pose estimation under 80 ms per frame, making it suitable for dynamic and large-scale environments.
- Enhanced Reconstruction Accuracy: Combines RGB-D input with robust depth handling techniques to produce high-fidelity 3D models.
- Versatility: Validated on both synthetic and real-world datasets, demonstrating reliable performance across diverse settings.
- Compatibility: Designed to work with standard RGB-D devices, including smartphones and consumer-grade depth sensors.