Matching of Satellite, Aerial and Close-range Images Luigi Barazzetti Politecnico di Milano, ABC Departement Lab. G@icarus, via Ponzio 31, Milan Lab. IC&T, Corso Premessi Sposi, Lecco luigi.barazzetti@polimi.it - http://www.icet-rilevamento.lecco.polimi.it Introduction to image matching Satellite, aerial and close-range data The image matching problem What stuff in the left image matches with stuff on the right? ‘Stuff’ … what does it mean? points, lines, areas, … ? Image matching 2 Introduction to image matching 3 Satellite, aerial and close-range data Image acquisition platforms (Photogrammetry and Remote Sensing) > 700 km Satellite images Aerial images UAVs Close-range images Waterproof equipment <1m Image matching Introduction to image matching 4 Satellite, aerial and close-range data Matching is easy! Why? Point of view Illumination Good texture Absence of occlusions … Figure by Noah Snavely Image matching Introduction to image matching 5 Satellite, aerial and close-range data Harder case! Why? An important consideration: human operators can easily extract corresponding points, objects, … Figure by Noah Snavely Image matching Introduzione al controllo statico delle strutture 6 Harder? (Nasa Mars Rover images) Figure by Noah Snavely Image matching Introduction to image matching 7 Satellite, aerial and close-range data Harder? SIFT key-points (automatic) Figure by Noah Snavely Image matching Introduction to image matching 8 Satellite, aerial and close-range data Applications Photogrammetry and Remote Sensing • fiducial mark measurement (target detection and matching) • tie point extraction • DTM/DSM generation • image registration • … Computer Vision • object recognition • multiple view analysis (3D reconstruction, panoramic images, HDR images, medical data alignment, …) • motion capture, object tracking • … … too many to be exhaustively listed Image matching Introduction to image matching Satellite, aerial and close-range data Motion capture: tracking a limited number of points Image matching 9 Introduction to image matching Satellite, aerial and close-range data 2.5D vs 3D models from aerial images Image matching 10 Introduction to image matching Satellite, aerial and close-range data Dense image matching and oblique imagery Image matching 11 Introduction to image matching Satellite, aerial and close-range data Example: Apple iOS Maps (the Cathedral of Milan) Image matching 12 Introduction to image matching 13 Satellite, aerial and close-range data Example: Apple iOS Maps The Statue of Liberty that was missing before now shows a 3D imagery of the landmark and the details of the structures surrounding it Image matching Introduction to image matching Satellite, aerial and close-range data Shaded and textured model Image matching 14 Introduction to image matching 15 Satellite, aerial and close-range data Point matching • Goal: extraction of good points • How can we define “good” and “bad” candidates? • Moving the window in any direction gives a big change “flat” region: “edge”: “corner”: no change in all no change along the significant change in all directions edge direction directions Slide adapted from Darya Frolova, Denis Simakov, and Noah Snavely Image matching Introduction to image matching 16 Satellite, aerial and close-range data Classification of point matching techniques Intensity-based matching: image data is used in form of grey values. Most prominent methods are cross-correlation and least squares matching (LS-matching). Also called "area-based" matching. Give sub-pixel accuracy, in extreme cases 1/100 pixel and better Feature-based matching: requires the extraction of basic image features, like blobs, corners, junctions, edges, etc. first. Matching is performed between these features. Features are sometimes more stable with regard to reflectance characteristics Relational matching: uses geometrical or other relations between features and structures (combination of features). Correspondence is established by tree-search techniques. These methods are not very accurate but usually robust Image matching Introduction to image matching 17 Satellite, aerial and close-range data Relational matching: uses geometrical or other relations between features and structures (combination of features). These methods are not very accurate but usually robust. They do not require good approximation Structural description: set of primitives and their inter-relationships Image matching Introduction to image matching 18 Satellite, aerial and close-range data Intensity-based matching: Normalized Cross-Correlation The solution (best match) is found at max(r(x, y)). r is limited to the region [-1, 1]. False or weak matches are indicated by small (r≤0,5). However, large r do not always indicate good, stable matches (e.g. in case of multiple solutions or in cases of weak signal in template and patch) r= Image matching Introduction to image matching Satellite, aerial and close-range data Feature-based matching: Foerstner operator Image matching 19 Introduction to image matching Satellite, aerial and close-range data Feature-based matching: modern operators New trend based on local features with detectors and descriptors Computer Vision methods: their use in Photogrammetry and RS is attractive Example with the SIFT operator: panoramic photography Brown, M. and Lowe, D.G., 2003. Recognizing Panoramas. International Conference on Computer Vision, 2: 1218-1225. Image matching Introduction to image matching Satellite, aerial and close-range data Feature-based matching: modern operators Picture with a Compact Camera (FOV) 50 x 35° Do you want a 360° reconstruction? Brown, M. and Lowe, D.G., 2003. Recognizing Panoramas. International Conference on Computer Vision, 2: 1218-1225. Image matching Introduction to image matching Satellite, aerial and close-range data Feature-based matching: modern operators Several pictures with a Compact Camera (FOV) Do you want a 360° reconstruction? Panoramic images (panoramas) Brown, M. and Lowe, D.G., 2003. Recognizing Panoramas. International Conference on Computer Vision, 2: 1218-1225. Image matching Introduction to image matching Satellite, aerial and close-range data Image matching plus constraints (geometry) Take pictures with a rotating camera Image matching Introduction to image matching Satellite, aerial and close-range data Feature-based matching: modern operators Extract SIFT features: detector / descriptor Geometrically invariant to similarity transforms Some robustness to affine change Automated methods outliers !!! Brown, M. and Lowe, D.G., 2003. Recognizing Panoramas. International Conference on Computer Vision, 2: 1218-1225. Image matching Introduction to image matching Satellite, aerial and close-range data Feature-based matching: modern operators Outlier rejection is fundamental RANSAC LS line Image matching Introduction to image matching Satellite, aerial and close-range data Feature-based matching: modern operators Outlier rejection is fundamental RANSAC RANSAC line Image matching Introduction to image matching Satellite, aerial and close-range data Feature-based matching: modern operators Outlier rejection is fundamental RANSAC (homography) Image matching Introduction to image matching Satellite, aerial and close-range data Feature-based matching: modern operators Final radiometric correction Extension to N images and final blending Brown, M. and Lowe, D.G., 2003. Recognizing Panoramas. International Conference on Computer Vision, 2: 1218-1225. Image matching Introduction to image matching Satellite, aerial and close-range data Feature-based matching and satellite images Similar approach but different operator (SURF) Input data over Las Vegas: • 1 ASTER image 5627x5001 pixels • 1 LANSAT TM image 7811x7011 pixels resolution 15 m resolution 30 m Image matching Introduction to image matching Satellite, aerial and close-range data Feature-based matching and satellite images Image 1 4141 interest points Image 2 14787 interest points Descriptor comparison 502 matches (outliers??) Image matching Introduction to image matching Satellite, aerial and close-range data Feature-based matching and satellite images Hypothesis: similarity transformation Xn Yn a b Xo c cos α = −b a * Yo + d S * − sin α sin α Xo DXo * + cos α Yo DYo Yn Y0 Yn Yn Y0 Y0 X0 Xn Xn X0 Translation Scaling Image matching Xn X0 Rotation Introduction to image matching Satellite, aerial and close-range data Feature-based matching and satellite images RANSAC: 395 inliers (from 502 matches) Final Least Squares adjustment Image matching Introduction to image matching Satellite, aerial and close-range data References Brown, M. and Lowe, D.G., 2003. Recognizing Panoramas. International Conference on Computer Vision, 2: 1218-1225. Brown, M. and Lowe, D.G., 2007. Automatic panoramic image stitching using invariant features. International Journal of Computer Vision, 74(1): 59-73. Gruen, A., 1985. Adaptative least squares correlation: a powerful image matching technique. South African Journal of Photogrammetry, Remote Sensing and Cartography, 14(3): 175-187. Gruen A., Geomatics in the 21th Century State of the art and future perspectives, Course at Politecnico di Torino Gruen A., Image Matching for DSM Generation, Compact Course at Politecnico di Milano Some slides and pictures from A. Gruen, M. Scaioni, S. Seitz, R. Szeliski, N. Snavely, M Brown, D. Lowe, D. Frolova, D. Simakov Image matching