antialiased cnns

Standard CNNs produce different feature activations for the same image shifted by even a single pixela property known as lack of translation equivariancewhich makes them fragile to small input perturbations and reduces the stability of learned representations.

The project introduces blur pooling as a simple, theoretically grounded fix that makes these operations equivariant to subpixel shifts.

The approach applies a low-pass filter (Gaussian blur) before any stride operation in the network, following the classical signal processing principle that downsampling should be preceded by anti-aliasing to prevent high-frequency aliasing artifacts.

This modification is compatible with existing CNN architectures (ResNet, DenseNet, VGG) and adds minimal computational overhead, while producing models that are more consistent across small input shifts and that often generalize better on classification and segmentation tasks.

The implementation is available as a drop-in replacement for standard pooling layers in PyTorch.

Computer vision researchers studying the theoretical properties of CNNs use antialiased convolutions as a building block when shift-equivariance is a meaningful constraint for their task.

Practitioners working on applications sensitive to precise spatial consistencymedical image analysis, satellite image processing, quality inspectionbenefit from the improved output stability.