Chapter 11: Structure from Motion and SLAM
Recovering 3D structure and camera motion from images, including camera calibration, pose estimation, bundle adjustment, and simultaneous localization and mapping.
Chapter Overview
How does a robot know where it is? How can we build a 3D model from photos? This chapter tackles the fundamental problem of recovering 3D structure and camera motion from 2D images.
What is this chapter about? We learn to calibrate cameras, estimate camera poses, optimize 3D reconstructions, and build maps in real-time for autonomous navigation.
Why does this matter? These techniques enable:
- Augmented reality: Placing virtual objects in real scenes requires knowing the camera pose
- Autonomous vehicles: Self-driving cars must understand 3D geometry
- Photogrammetry: Creating 3D models from drone or phone photos
- Robot navigation: SLAM lets robots explore unknown environments
How the topics connect: We start with camera calibration—measuring the intrinsic parameters of cameras. Pose estimation recovers where the camera is relative to known 3D points. Bundle adjustment jointly optimizes everything. SLAM does this in real-time as the robot explores.
Chapter Roadmap
Click any topic to jump in
Camera Calibration
Recovering intrinsic parameters (focal length, principal point, distortion) from checkerboard images — the prerequisite for all 3D geometry.
Pose Estimation
Determining camera position and orientation from known 3D-2D correspondences via PnP — essential for AR and localization.
Global optimization and real-time mapping
Bundle Adjustment
Joint optimization of all cameras and 3D points by minimizing reprojection error — the gold standard for accurate reconstruction.
SLAM
Simultaneous localization and mapping in real-time — building maps while tracking position for autonomous navigation.
Sign up to unlock this chapter
This chapter is part of PixelBank Premium. Create a free account, then upgrade to read the full lesson — concepts, walkthroughs, and exercises.