Suggestions on improving detections in low light conditions

[vishakha-lall]

I am working on AprilTag detections in a simulator environment which is dark and the (physical) tag is stuck on a screen. (See attached) The above image is captured when the lights in the simulator room are turned on. However, as a use case, I have frames that look like I want to know if anyone has tried some denoising/exposure enhancement techniques that do not affect the detections? I am able to do gamma adjustment and the tags are visible in the image but not detected. Things I have already tried

• Gamma Adjustment
• Histogram Equalisation
• Increase Brightness in PIL ImageEnhancer
• Adjusting decode_sharpening between 0.25-2 Is there anything that I can try in terms of preprocessing or modifying detection parameters/lowering thresholds?

FYI, the image size is 1920x1080 and the tag sizes are 30mm and 40mm

[brmarkus]

In a similiar challanging environment (incar cabin) infrared camera sensors are used. Would that be an option for you? Or print the tags on reflection capable paper?

[vishakha-lall]

@brmarkus thank you for your response, at this time probably not. I am trying to see if I can use some image processing techniques before I consider suggesting hardware changes. I am intrigued about reflection capable paper, would you be able to share an example?

[brmarkus]

Think about foils like on car number plates, or stripes on emergency vehicles, e.g. "https://www.reflecto.shop/".

With using infrared camera sensors the vehicle number plate recognition&detection got greatly improved (especially with bright light sources).

[brmarkus]

Isn't your environment static...? What do you use the tags for - in a AR/VR environment to dynamically place the various screens to composite the "plane's control board"?

What about detecting the table based on its shapes, contour, corners, edges, maybe apply perspective transformations? (like e.g. https://learnopencv.com/automatic-document-scanner-using-opencv/)