190 likes | 217 Vues
Submit your image to the search engine for building recognition. SIFT features ensure scale and rotation invariance, robustness to illumination changes, and accurate landmark identification. Explore the solution overview and demo scenarios for a reliable recognition system.
E N D
Aim of the project Take your image Submit it to the search engine Similar images are found Name of the building appears
Solution overview Image representation by SIFT keypoints Sky and hedge detection Non object features are removed Similarity matching of features with database
Properties of SIFT features Scale invariance Rotation invariance Robustness to illumination changes Robustness to reasonable perspective transformation
SIFT features computation Keypoint detection (local minima and maxima) Orientation of intensity gradients in the neighborhood Detection of significant orientations Histogram of orientations
Keypoint Refinement Sky detection (fast) Thresholding, morphological operations Applied to database and query images Hedge detection (slower) Hough transformation Closing morphology operation Applied to database images
Similarity of features Matching between keypoints Take only reliable matches Similarity: number of matched keypoints
Results SIFT Features outperformed PCA and Color Histogram Comparison Leave-one-out Validation First match correct: 95.5% First or second match correct: 98.5%