Exploiting redundancy for reliable aerial computer vision 1 Digital Image Increase 2
Images Worldwide 3 Terrestrial Image Acquisition 4
Aerial Photogrammetry 5 New Sensor Platforms Towards Fully Automatic Photogrammetric Reconstruction Using Digital Images Taken From UAVs. 6
Airborne vs. MAVs 7 Applications 8
Application: Virtual Habitat [Leberl et al. IEEE Computer 2010] 9 Application: Architecture and Cultural Heritage [Zebedin 2010], [Irschara 2010] 10
Application: Construction Site Monitoring 11 Application: Mining 12
Outline 3D Reconstruction Semantic Classification Discussion & Outlook 13 3D Reconstruction 14
Structure from Motion 15 Structure from Motion 16
Structure from Motion X 1 X 5 X 4 X X 2 X 3 X 7 minimize g Rt[K]X 6 R 1, t 1, [K 1 ] R 2, t 2, [K 2 ] R 3, t 3, [K 3 ] 17 Structure from Motion Images Pose Prior Calibration Pose Prior Feature Extraction Coarse Matching Detailed Matching Geometric Verification Geometric Estimation Local Descriptors Image Overlap Matches Epipolar Graph Camera poses 3D points 18
Aerial Photogrammetry 19 Global Depth Map Optimization ( ) ( ( ), ) ( ( ), ) ( ) 20
KEY view 21 DEPTH map ( ) ( ( ), ) min 22
DEPTH map ( ) ( ( ), ) min 23 24
Primal-Dual Optimization 25 Distributed Visual SLAM Low framerate, for map extension and relocalization Pose Full framerate, cheap features, for visual servoing Very little data needed, provide only data necessary for the current environment Maintain global map, use expensive features! Set of updateable, geo-referenced virtual cameras 26
Dense Reconstruction On-the-Fly 27 Dense Reconstruction On-the-Fly 28
Semantic Segmentation Motivation Semantic Interpretation 29 Building? Street? Tree? Water? 10 cm/pixel 30
Motivation a lot of overlapping images 3D information 31 Fusion - Model - Robust against outliers - Preserves sharp edges 32
Fusion - Model Some observations 33 Fusion Semantic Interpretation aggregated refined 34
Fusion - Color and Height Color: Wavelet-Fusion Height: TV-Fusion 35 Some observations in ortho-view 36
Evaluation Redundant Interpretation Remember evaluation on single images: Graz 89.5%, Dallas 92.5% 37 Large-Scale Results Dallas, 4 tiles, each 240 x 240 m, 15 cm Building Green Area Waterbody Tree Streetlayer 38
SF 10 x 3 tiles 2500 x 700 m 15 cm Building Green Area Waterbody Tree Streetlayer 39 Large-scale Results Graz, 7 km 2, 155 images, 20 x 10 tiles, 8 cm Building Green Area Waterbody Tree Streetlayer 40
Conclusions & Future Work semantic 3D model 41 Videos/Code/Papers see 42
Acknowledgments 43