Camera Exposure and Computer Vision

What is Camera Exposure?

Exposure is the amount of light accumulated at the image plane during a single capture, set by the lens aperture, the exposure time, and the scene's luminance.

Exposure Value (EV) is a related but distinct quantity: a logarithmic camera-setting scale, not the accumulated light itself. In the APEX system, EV = log₂(N²/t), where N is the F-number and t is the exposure time in seconds. The same EV yields different image-plane exposures as scene luminance changes. Sensitivity is not part of EV either: ISO enters the APEX equation separately, as the speed value Sv.

A series of gray patches is a common way to assess how an exposure rendered in a captured image. Exposure is not the "brightest bright" relative to the "darkest dark" of an image: that's dynamic range.

What is Exposure?

Correct exposure is important to all computer vision tasks as any over-exposure or under-exposure results in degradation of contrast and in extreme cases, a complete loss of information. Here is a picture of a simulated vehicle under various exposure values.

Even to the human observer, it is difficult to see details in over or under-exposed images.

What is Camera Exposure Value — Different Camera Exposures

Blasinski & Farrell, "Optimizing Image Acquisition Systems for Autonomous Driving"

Gain and ISO

Machine vision cameras usually have no ISO setting at all. They expose gain: analog gain applied to the pixel signal before the ADC, and digital gain applied to the digitized value afterward. The two are not equivalent, because where the gain acts relative to each noise source matters.

Analog gain amplifies the signal and its photon shot noise together, and because it acts before the ADC it shrinks the relative contribution of the noise added after the gain stage: ADC and quantization noise. Read noise accumulates across the reset, source-follower, column-amplifier, and ADC stages, so only the post-gain portion is suppressed; in a dim scene where those downstream terms dominate, a stop of analog gain can buy real SNR. Digital gain only rescales values that are already digitized, noise included, so it brightens the image without improving it.

Where a camera does offer ISO, treat it as a system-level rating rather than a raw gain knob: ISO 12232 defines digital-camera speed and exposure index from the processed output response, so two cameras at the same ISO number can apply different analog gains. ISO is what the APEX equation carries as the speed value Sv. Set the exposure time and the aperture first, add light to the scene if you can, and raise gain last, once a shorter integration time is the only way left to stop motion.

Exposure Has a Major Impact on Computer Vision and Machine Vision

Blasinski and Farrell demonstrated how the mean average precision of an SSD-MobileNet detector drops significantly with incorrect exposure. ResNet, the other network named in that work, is a classification backbone rather than a detector: it supplies features, and a detection head is built on top of it. Even after training with multiple exposure values, mean average precision dropped by 25 percentage points when over-exposed by four stops (four EV).

Computer Vision Degradation due to Exposure — Exposure and Machine Vision

References and Related Links Blasinski & Farrell, "Optimizing Image Acquisition Systems for Autonomous Driving"

Trying to Determine Your Camera Requirements?

Use our free web-based AoV Calculator to determine your system's Field of View Requirements. Then, use the M12 Lens calculator to match your requirements with the available lenses. Our Depth of Field Calculator also provides the hyperfocal distance and depth of field for every sensor and lens combination.

We also have a couple of other calculators that many engineers find interesting.