Sparse Checkerboard Corner Detection from Global Perspective

Detecting corners from an image is an essential step for camera calibration in geometric computer vision and image processing applications. In this paper, a novel framework is proposed to detect sparse checkerboard corners with a global context from an image. The proposed framework addresses two major problems that the previous neural network-based corner detection networks have had: locality and non-sparsity. Our framework encodes the global context from an image and uses the context to determine the per-patch existence of the checkerboard. It enables the network to distinguish between the checkerboard pattern and pattern-like noise in the image background while preserving pixel-level detection details. Also, the patch-wise sparse regularization is introduced using counting distribution to obtain clear-cut predictions while maintaining the true positive rate. The experimental results demonstrate that parsing the global context helps the proposed network to decrease false positive detection significantly. Also, the proposed counting regularization improves to detect true positives while decreasing false negatives concurrently. It enables the proposed network to precisely detect sparse checkerboard corners, leading to significant improvements over the state-of-the-art methods.