Evaluating edge detection? Don’t use RMSE, PSNR or SSIM.

Picture segmentation and edge detection are intently associated duties. Take this output from a coastal segmentation mannequin for instance:

The mannequin will classify each pixel as both land or ocean (segmentation masks). Then the shoreline is the pixels the place this classification modifications (edge map). Normally, edge detection might be performed utilizing the boundaries of the output of a picture segmentation mannequin.

I needed to make use of this relationship in my analysis to assist consider coastal picture segmentation fashions. Comparable analysis all use confusion matrix-based metrics like accuracy, precision and recall. These examine all pixels in a predicted segmentation masks to a floor reality masks.

The issue is these would possibly overestimate efficiency in an important area — the shoreline.

The vast majority of pixels are in the midst of the ocean or fully surrounded by land. This makes them simpler to categorise than these near the shoreline. You may see this in Determine 2. Sadly, these errors could also be shrouded within the sea of accurately categorized pixels.