1 / 22

Relevant UMass Research Edward Riseman Image Warping for Image Registration Howard Schultz

Presentations Computer Vision Research Lab. Relevant UMass Research Edward Riseman Image Warping for Image Registration Howard Schultz 3D Panoramic Imaging Howard Schultz 3D Terrain and Site Modeling Edward Riseman Aided Search and Target Cueing Gary Whitten

lajos
Télécharger la présentation

Relevant UMass Research Edward Riseman Image Warping for Image Registration Howard Schultz

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Presentations Computer Vision Research Lab • Relevant UMass Research Edward Riseman • Image Warping for Image Registration Howard Schultz • 3D Panoramic Imaging Howard Schultz • 3D Terrain and Site Modeling Edward Riseman • Aided Search and Target Cueing Gary Whitten • Knowledge-Based and Adaptive Image Processing Allen Hanson

  2. Knowledge-Based and Adaptive Image Processing Kollmorgen University of Massachusetts Amherst

  3. What? • Vision works reasonably well. • Vision doesn’t work. • Vision can be made to work better. Two premises and a conjecture

  4. A Simplified Example Algorithm for target (and algorithm!) cueing: Original Image Focus of Attention Area Process Area Enhance Potential Target Extract Horizon Visual Inspection ATR Algorithm Focus of Attention Area

  5. Hidden Assumptions and Constraints • Targets on or near horizon • Good contrast separation between water and sky • Targets large enough to detect as irregularities along horizon line • More or less calm seas • .......... LOTS OF VIOLATIONS!

  6. Violations Not enough contrast between water and sky Target below horizon Target too small Alternative Algorithm

  7. Ascender I System for Automatic Site Modelling Multiple Images Geometric Models • rooftop detection • line extraction • corner detection • perceptual grouping • epipolar matching • multi-image triangulation • geometric constraints • precise photogrammetry • extrusion to ground plane • texture mapping

  8. Ascender Works Detection Rates • Extensive evaluation • Good detection but high false alarms • Good geometric accuracy Ground Truth Ascender Results

  9. Ascender Doesn’t Work • Delivered to NEL (NIMA) • Applied to classified imagery • Performance not as expected • High failure rate (buildings not detected) • High false positive rate • WHY? Imagery didn’t No Buildings Detected! conform to design constraints!

  10. Stereo Terrain Reconstruction ISPRS Data Left Image Reconstruction Stereo Correlation Right Image

  11. Stereo Reconstruction Doesn’t Work No Correlations! Disparity Map Avenches Image - Right

  12. Making Vision Work • Algorithm Selection - which of many is the right one? • Function of data and explicitly represented constraints • Apply right algorithm at right time to right data • Parameter Selection • Automatically set critical parameters • Function of data and result of probes • Data Selection / Context Sensitivity • Data appropriate for algorithm? • Characterize data requirements

  13. Dealing with Context: Ascender II • Goals • 3D Geometric Site Model Reconstruction • Complex building structures • Context sensitive control strategies for applying algorithms • Multiple strategies • Multiple images • Multiple sensors EO, Digital Elevation Maps IFSAR, Multi/Hyperspectral

  14. Basic Principles • IU Algorithms work within correct context • constrained contexts • constrained object classes • Constraints • from domain knowledge, partial results, strategies • Of many strategies, only correct ones used • selective application • correct parameters • fuse results from individual strategies into complete reconstructions

  15. Ascender II Overview CP: Control Policy

  16. Schema/Belief Node Possible Values H1={j,..,k} Variable of interest H1 Evidence Policy H3 H2 First level of classification hierarchy H4 Second level (object subclass) Knowledge Representation • Combination of belief networks and • visual schemas, implemented in Hugin. • Encodes: Domain Knowledge • Acquired Site Knowledge • Control Mechanism • Allows both diagnostic and • causal inference. • Evidence policy defines how • IU strategies gather relevant • evidence according to context. • Hierarchical topology allows simpler • networks which reduces propagation • time and controls complexity.

  17. Preliminary Experiment • Goals • Show knowledge can improve accuracy of site model • Address NP-hard propagation using hierarchical nets • Approach • Site model from Ascender I with no filtering and loose polygon acceptance criteria => high detection rate but lots of errors • Classify Ascender I regions into {building, parking lot, open field, unknown} • Classify buildings into {multi-level building, single-level building} • Site model from Ascender II after classification

  18. 20 False Positive Buildings Degraded Ascender I Model No Knowledge and Loose Polygon Acceptance Criteria A 22 True Positive Buildings 1 Incorrect Model (A) (multi-level reconstructed as single level)

  19. First level of classification hierarchy Second level (object subclass) Bayes Net for Experiment Level 1 Building Open Field Parking Lot Complex Unknown Region Evidence Policy: none Level 2 Low Medium High L Junctions Line Count Planar Fit Good Bad Multi-Level Building Width Height Single Evidence Policy: none Evidence Policy: Evidence Policy: multi-EO: triangulateHeight() DEM: medianHeight() DEM: robustFit(plane) T Junctions Planar Fit Good Bad Yes No Evidence Policy: EO: junctions(T) <50 >50 Zero <5 >5 None Number T Junc. Contrast T Junc. Evidence Policy: Evidence Policy: EO: junction_count(T) EO: junction_contr(T)

  20. 43 Regions, 4 Misclassified, 1 Unknown 88% Accuracy Using Context Sensitive Strategies

  21. Ascender II Reconstruction 1 False Positive 23 True Positive Buildings 0 Incorrect Models (single, multi level) 3 New Functional Areas

  22. Conclusions • Vision works when properly constrained and focussed • Still a lot of work to do on: • Basic Algorithms • Knowledge Representations • Representing and Using Context and Constraints • Automatic Parameter Selection • Inferencing and Causal Reasoning • System Architectures

More Related