Rotation Matrix

A valid rotation matrix R is an orthogonal matrix, meaning its inverse is equal to its transpose: R⁻¹ = Rᵀ. This property ensures that applying the rotation and then its inverse returns a point to its original position. Furthermore, the determinant of a rotation matrix is always +1, which distinguishes it from other orthogonal matrices (like reflections, which have a determinant of -1) and guarantees it represents a pure, orientation-preserving rotation.

Successive rotations are combined through matrix multiplication. If R₁ rotates from frame A to B, and R₂ rotates from frame B to C, the composite rotation from A to C is R = R₂ * R₁. Crucially, matrix multiplication is not commutative; the order of rotations matters. For example, a 90° pitch followed by a 90° yaw yields a different final orientation than performing the yaw before the pitch. This property is essential for correctly chaining camera motions in SLAM or robotic kinematics.

Despite having 9 elements, a 3D rotation has only 3 Degrees of Freedom (DoF)—typically roll, pitch, and yaw. The 9 matrix elements are therefore highly constrained by 6 orthogonality conditions. This leads to alternative, more compact representations:

• Euler Angles: Intuitive but suffer from gimbal lock.
• Axis-Angle: A 3D vector whose direction is the axis and magnitude is the angle.
• Quaternions: A 4-element representation that avoids singularities and is computationally efficient for interpolation. The rotation matrix is often the final, stable representation used in calculations after converting from these other forms.

The inverse of a rotation matrix performs the exact opposite rotation. Due to orthogonality, computing the inverse is computationally cheap: it is simply the transpose of the matrix. This is pivotal in coordinate transformations. For a point p_w in the world frame and a camera with rotation R and translation t, the point in the camera frame is p_c = R * p_w + t. Conversely, to project from camera to world: p_w = Rᵀ * (p_c - t). This transpose operation is ubiquitous in pose estimation and 3D reconstruction.

In the pinhole camera model, the extrinsic parameters define the camera's pose in the world. They are composed of a rotation matrix R (3x3) and a translation vector t (3x1), combined into a single 3x4 matrix [R | t]. This matrix transforms 3D world coordinates (X, Y, Z) into 3D coordinates in the camera's reference frame before projection via the intrinsic matrix. Estimating R and t is the core problem of Perspective-n-Point (PnP) and Visual Odometry.

Estimated rotation matrices from noisy data (e.g., via SVD in essential matrix decomposition) may not perfectly satisfy orthogonality. A re-orthogonalization step, often using SVD again, is required to project the matrix onto the SO(3) manifold (the space of valid rotations). A common technique is to compute the SVD of the estimated matrix UΣVᵀ and set the valid rotation to R = U Vᵀ (ensuring det(R)=+1). This correction is critical in bundle adjustment and SLAM back-ends to maintain a consistent state.

Camera extrinsics define the position and orientation (pose) of a camera in a world coordinate system. They consist of two components:

• Rotation Matrix (R): A 3x3 orthogonal matrix defining the camera's orientation.
• Translation Vector (t): A 3D vector defining the camera's position. Together, they form the transformation [R | t] that maps 3D world points to the camera's coordinate system. Extrinsics are the primary output of pose estimation algorithms like PnP.

Six-Degree-of-Freedom (6DoF) describes the complete pose of a rigid body in 3D space, comprising three translational and three rotational degrees. For a camera:

• Translation: Movement along the X, Y, and Z axes (encoded in the translation vector).
• Rotation: Orientation around the X (roll), Y (pitch), and Z (yaw) axes (encoded in the rotation matrix). A full 6DoF pose is compactly represented by the combination of a rotation matrix and a translation vector.

Perspective-n-Point (PnP) is the classic problem of estimating a calibrated camera's 6DoF pose from a set of n known 3D points and their corresponding 2D projections in the image. The rotation matrix is a direct output of PnP solvers. Common algorithms include:

• EPnP: Efficient and accurate for n ≥ 4.
• Direct Linear Transform (DLT): A linear method.
• Iterative methods: Minimize reprojection error using non-linear optimization (e.g., Levenberg-Marquardt).

The Essential Matrix (E) is a 3x3 matrix that encodes the relative rotation and translation (up to scale) between two calibrated cameras. It is derived from epipolar geometry. Crucially, the rotation matrix between the two camera poses can be extracted from E via Singular Value Decomposition (SVD). The relationship is E = [t]_x * R, where [t]_x is the skew-symmetric matrix of the translation vector and R is the relative rotation matrix.

Singular Value Decomposition (SVD) is a fundamental matrix factorization used extensively in pose estimation to enforce geometric constraints on rotation matrices. Key applications include:

• Extracting R and t from the Essential Matrix.
• Solving the Orthogonal Procrustes Problem: Finding the optimal rotation matrix that aligns two sets of 3D points, which is central to algorithms like Iterative Closest Point (ICP).
• Ensuring orthogonality: Projecting a estimated 3x3 matrix onto the space of valid rotation matrices (SO(3)).

Quaternions are a four-dimensional number system ([w, x, y, z]) used as an alternative to rotation matrices for representing 3D orientation. They offer advantages for pose estimation:

• Compactness: 4 numbers vs. 9 for a matrix.
• No gimbal lock: Avoids singularities present in Euler angles.
• Efficient interpolation: Essential for smooth animation and SLAM. A quaternion can be converted to a 3x3 rotation matrix and vice-versa. Many optimization problems (like bundle adjustment) use quaternions internally to represent rotation.