I find it interesting that in small regions the consistency and nature of illumination matters very much at a perceptual level, but once we get to larger scenes there can be gross inconsistencies.
I hypothesize that illumination is perceptually relevant for determining object shape and reflectance, but once that is established illumination is more or less discarded perceptually as it is less useful for interacting with objects. It would also make sense then that certain complex objects that we are particularly well-adapted for perceiving, such as faces, would be perceived as a whole rather than just a collection of random objects and thus more sensitive to inconsistencies in illumination.
I was disappointed by the lack of quantitative results in the computational paper -- I know it's very hard to ground truth, but it was very unsatisfying to just have a lot of very pretty pictures and no real numbers (with the exception of training error from synthetic images).
I think that Ostrovsky et al. (2005) did a very good job at control experiements, but I still think that orientation of cubes are orthogonal to shading direction perception. It is because that the relative position of surface luminance is dependent both on cube orientation and the shading direction. And it is thru the relative position of surface luminance that we perceive the shading direction. I was thinking if they introduced variations to reflectance of cubes or sizes of the cubes the result might be as dramatic.
@Matt A - Agreed. It's especially disappointing because ground truth would not just have helped them report how good their results were, but seeing as they were training a classifier, they could have used as a feedback metric.
I also think there are at least a couple reasonable ways to go about measuring ground truth. You could, for example, take pictures of your scene with really really diffuse lighting to approximate a reflectance intrinsic. Then you could spray-paint everything matte white and take another picture to get the illumination intrinsic. Might take longer to generate a large training set, but at least you'd have a good one.
You could also generate a large training set synthetically by using a database of 3d models and using a physically based renderer to render them.
Oops. I realized that my comment above is missing two "not"s. What I actually intended was,
I think that Ostrovsky et al. (2005) did a very good job at control experiements, but I still think that orientation of cubes are *not* orthogonal to shading direction perception. It is because that the relative position of surface luminance is dependent both on cube orientation and the shading direction. And it is thru the relative position of surface luminance that we perceive the shading direction. I was thinking if they introduced variations to reflectance of cubes or sizes of the cubes the result might *not* be as dramatic.
I find it interesting that in small regions the consistency and nature of illumination matters very much at a perceptual level, but once we get to larger scenes there can be gross inconsistencies.
ReplyDeleteI hypothesize that illumination is perceptually relevant for determining object shape and reflectance, but once that is established illumination is more or less discarded perceptually as it is less useful for interacting with objects. It would also make sense then that certain complex objects that we are particularly well-adapted for perceiving, such as faces, would be perceived as a whole rather than just a collection of random objects and thus more sensitive to inconsistencies in illumination.
I was disappointed by the lack of quantitative results in the computational paper -- I know it's very hard to ground truth, but it was very unsatisfying to just have a lot of very pretty pictures and no real numbers (with the exception of training error from synthetic images).
ReplyDeleteI think that Ostrovsky et al. (2005) did a very good job at control experiements, but I still think that orientation of cubes are orthogonal to shading direction perception. It is because that the relative position of surface luminance is dependent both on cube orientation and the shading direction. And it is thru the relative position of surface luminance that we perceive the shading direction. I was thinking if they introduced variations to reflectance of cubes or sizes of the cubes the result might be as dramatic.
ReplyDelete@Matt A - Agreed. It's especially disappointing because ground truth would not just have helped them report how good their results were, but seeing as they were training a classifier, they could have used as a feedback metric.
ReplyDeleteI also think there are at least a couple reasonable ways to go about measuring ground truth. You could, for example, take pictures of your scene with really really diffuse lighting to approximate a reflectance intrinsic. Then you could spray-paint everything matte white and take another picture to get the illumination intrinsic. Might take longer to generate a large training set, but at least you'd have a good one.
You could also generate a large training set synthetically by using a database of 3d models and using a physically based renderer to render them.
Oops. I realized that my comment above is missing two "not"s. What I actually intended was,
ReplyDeleteI think that Ostrovsky et al. (2005) did a very good job at control experiements, but I still think that orientation of cubes are *not* orthogonal to shading direction perception. It is because that the relative position of surface luminance is dependent both on cube orientation and the shading direction. And it is thru the relative position of surface luminance that we perceive the shading direction. I was thinking if they introduced variations to reflectance of cubes or sizes of the cubes the result might *not* be as dramatic.