HELF-SLAM is a lightweight and efficient visual SLAM system that combines a hybrid learned frontend with a classical optimization backend.
It achieves real-time performance on a single GPU with low memory usage while maintaining robustness under challenging conditions such as low texture, illumination changes, and motion blur.
- Real-time SLAM with an efficient learned frontend
- Robust tracking using adaptive keypoint sampling and learned features
- Low GPU memory usage (~3 GB), suitable for mobile or edge devices
- Optimized backend with bundle adjustment
- Supports standard benchmarks such as TUM, EuRoC, and KITTI
- Runs at 30 FPS on a laptop GPU
The main implementation will be released soon.
Code coming soon.