SLAM-RS

Trying to implement a SLAM system in Rust.

Dependencies

• OpenCV
• nalgebra
• serde
• rerun (optional, for 3D visualization)
• Clap (for command-line argument parsing)
• tch-rs (optional, for deep learning depth estimation)

Usage

Build with Cargo:

cargo build --release

Run visual odometry example:

# Use default KITTI intrinsics
cargo run --example visual_odometry /path/to/video.mp4

# Specify custom camera intrinsics
cargo run --example visual_odometry /path/to/video.mp4 -- --fx 500 --fy 500 --cx 320 --cy 240

Run point cloud generation with real-time 3D visualization (Rerun):

# With Rerun 3D viewer (shows map, trajectory, matches, video in real-time!)
cargo run --example point_cloud --features rerun /path/to/video.mp4 -- --rerun

# Or save to PLY file (default, no Rerun)
cargo run --example point_cloud /path/to/video.mp4 -- --save-ply

# With custom camera intrinsics
cargo run --example point_cloud --features rerun /path/to/video.mp4 -- --rerun --fx 718.856 --fy 718.856 --cx 607.1928 --cy 185.2157

Run feature detection visualization:

cargo run --example visualize_features /path/to/video.mp4

Run monocular depth estimation:

# Single image
cargo run --example depth_estimation --features depth -- test.jpg

# Video with Rerun visualization (requires model files in weights/)
cargo run --example depth_estimation --features depth,rerun -- test.mp4 --cuda --rerun

# See docs/Deep-Learning.md for model setup

Modules

• feature: ORB feature detection and matching
• odometry: Camera intrinsics, pose estimation, trajectory tracking
• mapping: Keyframe selection, 3D point triangulation, map points, bundle adjustment
• depth: Monocular depth estimation with MonoDepth2 (optional, requires depth feature)

Examples

• visualize_features: Real-time feature detection and matching visualization
• visual_odometry: Full VO pipeline with trajectory tracking and visualization
• point_cloud: 3D point cloud reconstruction with triangulation and bundle adjustment
• bundle_adjustment: Demonstration of pose and point optimization
• depth_estimation: Monocular depth estimation with MonoDepth2 (requires --features depth)

See TODO for development status and docs/Deep-Learning.md for depth estimation setup.

FAQ

• Why is the map sparse?
  • This is a feature-based system using ORB features (~1000-2000 per frame). We only triangulate at corner-like features, not every pixel. This is similar to ORB-SLAM.
• How can I make the map denser?
  • Use the MonoDepth2 depth estimation module (see docs/Deep-Learning.md)
  • For dense reconstruction, you need:
    • Dense/semi-dense tracking (all high-gradient pixels)
    • Depth estimation/fusion
    • More computational resources

Features

Feature	Status	Cargo Flag
Feature-based VO	✅	(default)
Bundle Adjustment	✅	(default)
Rerun Visualization	✅	--features rerun
Depth Estimation	✅	--features depth
Loop Closure	🚧	-