CS 766 Final
Samareh Shahmohammadi
Automated Point Selection
To increase accuracy, it's good to use more points. However, it is quite time consuming to do this for any set of images. This led to my attempts at automating this process.
My goal was to produce that would yield results like those produced by hand selection. However, each seemed to have its own shortcoming.
Random Point Selection
1
It is unlikely that the combination of points across the series of exposures would provide enough data points to recover the camera response curve well enough.
Points selected evenly across full range from any one photograph
2
All images are subject to the limitations of the low dynamic range. Selecting points evenly across the range in one image will likely lead to redundancies, selecting points around one end of the irradiance spectrum.
​
For example, if an image is underexposed, most of the points selected will be chosen around the brighter areas, and ignore the lower range.
Points selected evenly across full range from mean of photographs
3
This approaches attempts to overcome the shortcomings of selecting point from any one photograph. It will certainly select points from the extremes.
​
However, if most of the images are under- or over-exposed, the same issues of concentrating the selected points around one end of the range.
My Solution
After trying multiple things, I chose to tackle the issue in the following way:
​
Split the points in half. For the first half select points that lie evenly across the full intensity range from what I term the “middle photograph”. This is the one that has a mean of intensities closest to the average of the highest and lowest means. In other words:
-
Find the mean of intensities for all images​
-
Take the average A of the highest and lowest means
-
Select the image that has the mean of intensities closes to that average A
This is a variation on approach 2, but it and it bypasses issue of too many over- or under-exposed images.
​
This does not necessarily choose points at the extremes, where most points are dark or light. That's why for the second half of the points, I do the same on the mean of the photographs (as in approach 3). This ensures that points are selected on the extremes across all images.
Below are the general and local tonemapped images based. If you look closely, they toggle between images based on hand-picked points and automatically selected points; this is or sake of comparison. The slightly lighter ones are the ones based on points chosen by hand. As you can see, the difference is not too great with this approach.
​
