Describe the feature and motivation

The cv::barcode::BarcodeDetector relies on hardcoded parameters in bardetect.cpp.

The barcode localisation algorithm is very good, but it has some limitations in terms of high-res and small barcode inputs that make it both resolution and barcode scale dependent.

I had managed to figure out four values that impact this:

• the input image is downsampled to 512px by default (a workaround is to use np.lib.stride_tricks.sliding_window_view to process the image in patches, and NMS the resulting boxes afterwards which is inefficient).
• the gradient magnitude threshold is permanently set to 64.
• window sizes for filters have been empirically determined with respect to input size and is not possible to adjust them {0.01f, 0.03f, 0.06f, 0.08f} for other scenarios.
• the NMS Box threshold is calculated and set for a 512px wide image

There are a lot more parameters in bardetect.cpp but these are the most relevant ones which need to be tuned to improve the barcode localisation performance.

This is what the stock detector returns for a 2048x1536 input image.

Obviously this is very poor performance. But after manually adjusting these parameters, I am able to successfully localise each barcode in the image.

Additional context

Proposal on how to set detector parameters, and reference values used for the test pattern.

detector = cv2.barcode.BarcodeDetector()

div = float(np.max(frame.shape)) / 512.
detector.setDownSampleThresh( div * 512 ) # prevent downsample
detector.setDetectorScales( [ 0.01/div, 0.03/div, 0.06/div, 0.08/div ] )
detector.setGradientThresh( 256 )

This should keep compatibility and maintain the current behaviour of the standard barcode detector.

Reference image

LE: Updated proposal to set detector parameters per 24903