Distinguishing shadows from surface boundaries using local achromatic cues.

Area of Science:

• Computer Vision
• Computational Neuroscience
• Image Processing

Background:

• Accurate visual scene parsing requires distinguishing surface boundaries from shadow edges.
• While chromatic cues are studied, achromatic cues (contrast, texture, blur) for shadow detection are less understood.

Purpose of the Study:

• To analyze image properties distinguishing achromatic shadow edges from occlusion edges.
• To understand the role of contrast, texture, and penumbral blur in edge classification.

Main Methods:

• Developed and analyzed a large database of hand-labeled achromatic shadow and occlusion edges.
• Utilized logistic regression on a Gabor Filter Bank (GFB) model and a Filter-Rectify Filter (FRF) neural network.
• Compared machine classifier performance (GFB, FRF) with human edge classification accuracy.

Main Results:

• Both highest and lowest contrast edges were more likely to be occlusions than shadows.
• Contrast cues alone achieved nearly 70% accuracy in distinguishing edge types.
• GFB and FRF models achieved >80% accuracy, with FRF showing better texture sensitivity and correlation with human performance.

Conclusions:

• Achromatic cues, particularly contrast and texture, are crucial for differentiating shadow edges from surface boundaries.
• Machine learning models (GFB, FRF) can effectively classify these edges, mirroring human visual system capabilities.
• The FRF model's stronger agreement with human performance suggests texture plays a significant role in human shadow perception.