Camera Pose Estimation Using Viewpoint-Based Image Clustering

The patent segments the image retrieval process by introducing viewpoint data (azimuth and elevation angles) as a separate filtering criterion. Instead of relying solely on image similarity algorithms, the patent divides the retrieval process into two stages: first filtering images based on viewpoint distance thresholds, then performing detailed similarity comparison only on pre-filtered candidates. This segmentation reduces the number of images requiring computationally expensive similarity analysis.

The patent performs preliminary filtering of candidate images using viewpoint data before executing the full image retrieval algorithm. By calculating azimuth and elevation distances in advance and applying threshold-based pre-filtering, the system prepares a reduced set of candidate images that are more likely to be truly similar. This preliminary action prevents the retrieval algorithm from wasting time on obviously dissimilar images.

The patent applies local quality by using viewpoint data to selectively process only those images that have similar viewing angles to the query image. Instead of attempting feature extraction from all retrieved images, the system identifies and processes only the local subset of images whose viewpoint distances (azimuth and elevation) fall within predefined thresholds. This localized processing ensures that feature extraction is performed on images with comparable scene geometry and object configurations.

The patent changes the parameters used for image selection from purely similarity-based metrics to include viewpoint parameters (azimuth angle difference and elevation angle difference). By introducing these angular parameters and setting threshold values, the system transforms the retrieval criterion to better suit dynamic scenes with multiple objects, where viewpoint consistency is more important than overall image similarity.