CameraVQ: Vector-Quantized Representations for Monocular Camera Calibration

초록

We present a monocular camera calibration method, dubbed CameraVQ. It reformulates monocular camera calibration as classification over vector-quantized camera intrinsics. Existing methods based on geometric cues or direct regression often suffer from unstable optimization and poor generalization. In contrast, CameraVQ learns a discrete codebook of camera intrinsics and predicts the latent code from a single image. This discrete formulation constrains predictions to a statistically learned manifold of valid configurations, enabling robust calibration and strong generalization. Through extensive evaluation on diverse calibration benchmarks, CameraVQ achieves state-of-the-art performance on the majority of datasets.

키워드