Ground truth is Mask R-CNN's dog mask on this photograph, its highest-scoring instance at 0.999; the prediction starts as a perfect copy and you damage it. Everything here is counted on the mask's own 480 × 322 grid, a 2.6× downsample of the 1264 × 848 photograph, so n = 10 px erodes about 26 pixels of the original. The dog is only 4.67% of the frame, so a model that labels every pixel background is right about 95.3% of them — press all background and read the accuracy off. IoU divides the agreement by the union of the two dog regions, and the same prediction scores zero. Predict before you drag: at the far end of the slider, which of the three degradations is worst? Pixel accuracy ranks them dilate, shift, erode — worst first — and IoU ranks them erode, shift, dilate. The two metrics reverse the order completely.