Checkerboard Corner Localization Accelerated with Deep False Detection for Multi-camera Calibration

Camera calibration is an indispensable step in the fields of robotics and computer vision, which includes 3D reconstruction and camera motion estimation. Before camera calibration, detecting image features and matching their correspondence are necessary to understand the structure of the world from multiple images. For an accurate multi-view camera calibration, calibration object, such as checkerboard, is used. When using checkerboard, handcrafted feature methods precisely detects checkerboard corners but are slow in multi-view camera setting due to inclusion of cornerless images. Conversely, neural network-based detectors can quickly detect corners in images. Even though neural network-based detectors are known for its localization of corners in noiseful images, such as motion blur, accuracy of noisy corner is still questionable. Therefore, in this paper, we propose a novel framework for multi-view camera system to distinguish reliable image corners to be used in the calibration via neural network and sub-pixel refinement estimated from the handcrafted refinement method. Through a neural network-based detector, the framework detects corners while removing seemingly cornerless images with speed. The handcrafted refinement method accurately refines the sub-pixel location of detected corners. Hence, the proposed framework localizes checkerboard corners faster and more accurately than solely using handcrafted or state-of-the-art neural network-based methods.