Spatial Pyramid Matching for Scene Classification

After get the filter response of the scenes, based on k-means, using a dictionary to represent the scenes. On this basis, applying spatial pyramid matching to include information of spatial structure of the scenes. Then according to the distances among the training scenes and the testing scene, realizing scene classification.

Keypoints – Detectors, Descriptors and Matching

Extracting the interest points from the images, and using BRIEF feature descriptors to match them (find correspondences) between images. Keypoints are found by applying the Difference of Gaussian (DoG) detector. And based on the principal curvature ratio in a local neighborhood of a point, removing the edge points which are not desirable for feature extraction. Then generate a static set of test pairs based on uniform distribution. Finally get the BRIEF descriptors and the matching between images.



Homographies & RANSAC

Based on matched points to compute the 3 × 3 transformation matrix and applying coordinate normalization to realize the planar homography. Then stitch two views together to get the final mosaic view. For the models with outliers, use RANSAC to get robust inliers by iteration.