antialiased-cnns

Why we included this project

A one-pixel shift in the input can flip a convolutional network's prediction, a real headache for image classification and object recognition. This PyTorch package fixes that with the BlurPool layer, a direct replacement for stride-2 max and average pooling that blurs before downsampling and restores shift invariance without adding parameters. If you already use torchvision backbones, the big convenience is the pretrained antialiased weights for common architectures like AlexNet, VGG, ResNet, DenseNet, and MobileNetV2, so you can load a checkpoint instead of training from scratch. You can also apply BlurPool to your own architecture and copy existing weights over to continue fine-tuning, which keeps the upgrade path simple. The README reports measured ImageNet gains in both top-1 accuracy and prediction consistency across shifted inputs, so you know what to expect before adopting it.