Awesome

Camera parameters for Human3.6M dataset

This repository contains camera parameters (intrinsics and extrinsics) for Human3.6M dataset in handy format. File camera-parameters.json stores all parameters extracted from metadata.npy (see metadata.xml from official code package) using generate.py script.

Camera parameters

• R - 3x3 rotation matrix
• t - 3x1 translation vector
• calibration_matrix - 3x3 calibration matrix
• distortion - 5 values [k1, k2, p1, p2, k3] (see this OpenCV article)

Projection (without distortion)

To build projection matrix P just multiply calibration and extrinsics matricies:

P = calibration_matrix @ np.hstack([R, t])
print(P.shape)  # (3, 4)

To project 3D point X to image plane:

X = np.random.rand(3)  # random 3D point
X_homo = np.hstack([X, 1.0])  # convert to homogenous

x_homo = P @ X_homo  # project
x = x_homo[:2] / x_homo[2]  # convert back to cartesian (check that x_homo[2] > 0)