Edge Detection Evaluation in Boundary Detection Framework Feng Ge Computer Science and Engineering, USC
Edge detection Detect pixels with strong gradient of “gray-level” Error – False negative(Missing ): Not detected Edges – False positive: detected false edges – Orientation error: shift from real position – Dislocation error: shift from real direction How to evaluate these errors? Edge Detection Error
Evaluation Criteria Ground Truth – Human or predefined results? Quantificaition – Measuring and expressing in number means good. Generality – Real images in large number Combined 3 criteria are good evaluation methods!
Overview Subjective vs Objective –H uman vision checking – Quantitative measurement With ground truth vs Without – Standard for evaluation – Some characters,e.g, continuation,coherence. Synthetic vs Real images – Simple structure – Complicated structures
Motive— in boundary detection framework Problem: Boundary detection algorithms work well in synthetic data, while poorly in real images This gap,we believe, is largely introduced by edge detection
Experiment Settings : Image Database Large: 1030 images Generality Unambiguous Manually extracted ground truth
Experiment Settings : Evaluation Flowchart
Experiment Settings : Detectors Edge & Line Detector: Canny & Line Approximation Boundary detector: Ratio-Contour
Experiment Settings : performance measurement
Experiment Original images image->edge->fragments->bounday->evaluation Synthetic images texture images->fragments --->bounday->evaluation ground truth->adding noise Semi-synthetic images original images->background -->bounday->evaluation ground truth->adding noise
Experiment --Synthetic images Result – Much better than original images Problem – Background correlation changed – Irregular background in texture images
Experiment – Semi-synthetic images Model simulation Edge-map error analysis
Result-1 Simulate edge missing Procedure: Sample ground truth, random delete some percentage of fragments
Result-2 Simulate edge detection error: missing & dislocation – Fix dislocation error, vary missing rate (a) – Fix missing error, vary dislocation error (b) (a) (b)
Conclusion Our noise model is close to real edge error, as regarding to the simulated result Edge missing and dislocation are mainly encountered errors in edge detection. Edge dislocation is more crucial in edge error compared with missing error
Discussion-1 Error introduced by line detection
Discussion-2 Model error – Gaussian distribution assumption Based on boundary detection – Globally, not locally – Introduce some error, but statistically, reasonable Image database – Low resolution – Ground truth error
Future work Distinguish errors introduced by line approximation from edge detection Noise model refinement Substitute line with curve in edge-map approximation Data base improvement
Thank You !