Segmentation Lab

Dilated (atrous) convolution: how far does one output pixel see?

Stacked 3×3 layers

Dilation rate

Each shaded cell is an input pixel that influences the output pixel in the centre (outlined); darker = more paths through the layers. A dilated 3×3 kernel has 9 taps spaced rate pixels apart, so it covers a wider area with the same 9 weights.

Measuring a segmentation: pixel accuracy vs. IoU

correct (true positive)false alarm (false positive)missed object (false negative)

Predicted mask width

Predicted mask height

Drag the predicted (blue/orange) mask. The image is 20 × 20 = 400 pixels; the object is the dashed region. IoU = TP / (TP + FP + FN), Dice = 2·TP / (2·TP + FP + FN), pixel accuracy = correct pixels / all pixels (background included).