I am surprised by the idea of adaptive windows being inherent to our visual model of estimating ATF. Is there a suggested biological method for performing that sort of process? A level of hierarchical processing of data with only one level more activated than others? Some sort of bayesian estimation of likelihood of different atmosphere estimates done in parallel?
The painted polyhedra is an interesting model for estimating 3d shape to disambiguate reflectance and illuminance changes. However, it assumes no obscurance, which is very limitted. It would be interesting to present people with a 3d occluded view and figure out how people (or monkeys or fish) move their viewpoint to disambiguate them. Given that birds are able to estimate stereo by moving their heads rapidly, I suspect there is at least some stateful estimation from multiple viewpoints in humans.
I agree with Matt about occlusion being interesting. Also, the examples presented by Sinha and Adelson do not show gradient intensity patterns. It might be interesting as well running the algorithm in these cases, just by looking locally at junctions. Gradients tend to be smooth, so the little change in intensity around junctions might be safely ignored during the brightness equalization process. I can imagine a particular case where this kind of situation might generate confusion in the model, though.
With regards to Matt's first question, I find it particularly interesting that the adaptive window is not based not just on low-level scene organization like co-location, but can actually change shape based on higher-level percepts of scene geometry. For example the T-junction illusion in figure 18 (Adelson 2000) suggests that the windows used are the entire horizontal stripes.
I'd also posit that there are not just unrelated non-overlapping windows, but that different windows can occupy the same space and inform the brightness/lightness estimates of each other. For example, in Figure 24.17 you can get look at windows corresponding to the background circles or the foreground transparent squares. And, they are not unrelated- if you were to raise the luminance of one of the circles it would be perceived as brighter than the others.
What really interested me (in Sun & Perona's) study was how when they did the study on left-handed and right-handed people and both perferred left-sided lightness (and right-handed people were stronger). Though it seems that throughout history that everyone preferred the left (that we generally live in a "right-handed" world). These types of studies always lead me to question "what if" we had a bubble to "trap" newborns in and completely reverse all the lighting situations. (Which would seem to completely answer what effects the world and experience in handedness have). Would we completely skew their natural tendencies (like the parent who ended up teaching their child Klingon as their first language) or regardless would we just always be on the left.
I am surprised by the idea of adaptive windows being inherent to our visual model of estimating ATF. Is there a suggested biological method for performing that sort of process? A level of hierarchical processing of data with only one level more activated than others? Some sort of bayesian estimation of likelihood of different atmosphere estimates done in parallel?
ReplyDeleteThe painted polyhedra is an interesting model for estimating 3d shape to disambiguate reflectance and illuminance changes. However, it assumes no obscurance, which is very limitted. It would be interesting to present people with a 3d occluded view and figure out how people (or monkeys or fish) move their viewpoint to disambiguate them. Given that birds are able to estimate stereo by moving their heads rapidly, I suspect there is at least some stateful estimation from multiple viewpoints in humans.
I agree with Matt about occlusion being interesting. Also, the examples presented by Sinha and Adelson do not show gradient intensity patterns. It might be interesting as well running the algorithm in these cases, just by looking locally at junctions. Gradients tend to be smooth, so the little change in intensity around junctions might be safely ignored during the brightness equalization process. I can imagine a particular case where this kind of situation might generate confusion in the model, though.
ReplyDeleteWith regards to Matt's first question, I find it particularly interesting that the adaptive window is not based not just on low-level scene organization like co-location, but can actually change shape based on higher-level percepts of scene geometry. For example the T-junction illusion in figure 18 (Adelson 2000) suggests that the windows used are the entire horizontal stripes.
ReplyDeleteI'd also posit that there are not just unrelated non-overlapping windows, but that different windows can occupy the same space and inform the brightness/lightness estimates of each other. For example, in Figure 24.17 you can get look at windows corresponding to the background circles or the foreground transparent squares. And, they are not unrelated- if you were to raise the luminance of one of the circles it would be perceived as brighter than the others.
What really interested me (in Sun & Perona's) study was how when they did the study on left-handed and right-handed people and both perferred left-sided lightness (and right-handed people were stronger). Though it seems that throughout history that everyone preferred the left (that we generally live in a "right-handed" world). These types of studies always lead me to question "what if" we had a bubble to "trap" newborns in and completely reverse all the lighting situations. (Which would seem to completely answer what effects the world and experience in handedness have). Would we completely skew their natural tendencies (like the parent who ended up teaching their child Klingon as their first language) or regardless would we just always be on the left.
ReplyDelete