Texture perception Lavanya Sharan February 23rd, 2011
Typical texture perception display Image source: Landy & Graham (2004)
Typical texture perception display Image source: Landy & Graham (2004)
Typical texture perception display Image source: Landy & Graham (2004)
Typical texture perception display Image source: VPfaCGP Fig 8.5
Typical texture perception display Image source: VPfaCGP Fig 8.5 These are examples of texture segregation/segmentation tasks. Similar tasks in visual search (e.g., find a T among Ls)
Explaining performance at these tasks The ‘back pocket’ model. Image source: Landy & Graham (2004)
‘Back pocket’/LNL/FRF/Second-order model etc. Image source: Landy & Graham (2004) Input After 1st stage After 2nd stage Output
‘Back pocket’/LNL/FRF/Second-order model etc. Image source: Landy & Graham (2004) Lots of work on these models. Not tied to specific features (e.g., line terminations). Explain performance on many texture segregation tasks. Biological plausibility.
For example, Malik & Perona (1990) Image source: VPfaCGP Fig 8.3
For example, Bergen & Adelson (1988)
Back pocket model works on most lab stimuli. Image source: Ben-Shahar 2006
An failure case for the back pocket model. Image source: Ben-Shahar 2006 Text Textures Manual annotations The orientation gradient is negligible across the perceptually salient boundaries.
Lab stimuli vs. Real world stimuli Image source: Landy & Graham (2004)Image source: VPfaCGP Fig 8.2 Lots of psychophysics. Many computational models of perception. Hardly any psychophysics. Very few computational models of perception (mostly in computer vision).
Modeling texture appearance (Portilla & Simoncelli 2001) Like Heeger & Bergen, impose constraints iteratively. Four classes of constraints. Each set adds something about real world texture appearance. Analytical model (as opposed to patch-based models) allows a framework for understanding texture perception.
Modeling texture appearance (Portilla & Simoncelli 2001) Like Heeger & Bergen, impose constraints iteratively. Four classes of constraints. Each set adds something about real world texture appearance. Analytical model (as opposed to patch-based models) allows a framework for understanding texture perception.
Shape from texture Under assumption of isotropic texture patterns, one can estimate slant and tilt of surfaces. Image source: VPfaCGP Fig 8.7
Slant, tilt & perspective interact to produce texture distortions Image source: Todd et al. 2005
What about real world images? Torralba & Oliva (2002)
Summary ✓ Most perceptual studies think of texture as black-and-white simple shapes. ✓ We have learnt a lot from these stimuli. ✓ Time to examine real-world textures. Some methods to manipulate these exist (e.g., computer vision methods). ✓ Real world texture overlaps with real world materials. More next time.