Image Orientation Detection Based on Camera-subject Distances by Adaptive Control of Conditional Probability

In this paper, we propose a method of image orientation detection based on camera-subject distances. This method is characterized as an adaptive control of probabilistic decision by subject distances. The control is operated by switching of conditional probabilities of visual features at each subject distance category. The subject distance can be extracted from Exif. Evaluation and comparison is performed on 1681 consumer images. The experimental results show accuracy improvements on all visual feature combinations with negligible computational cost.