Is there a metric to measure the robustness of an application in different lighting conditions?

asked 2019-02-20 22:22:22 -0600

SRG-FTL gravatar image

updated 2019-02-20 22:42:23 -0600

Hello, I would like to know if there is a metric to measure the robustness to changes in lighting or different sources of light such as sun at sunrise, sun at noon, only at dusk, incandescent focus, low consumption focus, etc. That is, if you want to test a computer vision application in different lighting conditions such as those mentioned above, how could you measure that robustness in the application? Specifically, if I am doing object detection and want to test if my application still works in different lighting conditions. I do not know there is any unit of measurement in the international system or any metric that is used exclusively in computer vision. Thanks in advance.

edit retag flag offensive close merge delete



These will be statistical measures. You can measure real/false positives and negatives; precision and recall; error rate... Study them to decide for the most relevant for your application.

You can try global measures (using a big database of all the images in different conditions) or group the images for the conditions (you'll get measures for optimal conditions (noon), for more difficult conditions and most difficult conditions (backlight, fog, etc).

kbarni gravatar imagekbarni ( 2019-02-21 03:42:41 -0600 )edit

For a moment I thought it would be good to use some units of measurements such as lux or lumen but I do not know if that is correct because the idea that occurred to me was to say that my application was tested in conditions of "x" lumens and works correctly. That is, give results that my application works well in certain conditions, as well as microprocessor manuals that say that the maximum operating voltage is 1.8v or 3.2v for example. Do you think that this idea is correct for the field of computer vision?

SRG-FTL gravatar imageSRG-FTL ( 2019-02-21 20:17:35 -0600 )edit

It's not the light intensity that will cause problems, but the conditions. Imagine a sunset: your algorithm won't have the same performance if the sun is behind you (it should work well) or facing the camera (it will have worse results), even if the luminosity is the same.

I suggest to test your method in every condition your application can encounter to see the robustness.

kbarni gravatar imagekbarni ( 2019-02-22 12:35:09 -0600 )edit

@kbarni Thanks

SRG-FTL gravatar imageSRG-FTL ( 2019-03-18 00:50:48 -0600 )edit