Richard Zhang
Modern convolutional networks are not shift-invariant, as small input shifts or translations can cause drastic changes in the output. Commonly used downsampling methods, such as max-pooling, strided-convolution, and average-pooling, ignore the sampling theorem. The well-known signal processing fix is anti-aliasing by low-pass filtering before downsampling. However, simply inserting this module into deep networks degrades performance; as a result, it is seldomly used today. We show that when integrated correctly, it is compatible with existing architectural components, such as max-pooling and strided-convolution. We observe \textit{increased accuracy} in ImageNet classification, across several commonly-used architectures, such as ResNet, DenseNet, and MobileNet, indicating effective regularization. Furthermore, we observe \textit{better generalization}, in terms of stability and robustness to input corruptions. Our results demonstrate that this classical signal processing technique has been undeservingly overlooked in modern deep networks. Code and anti-aliased versions of popular networks are available at https://richzhang.github.io/antialiased-cnns/ .
| Task | Dataset | Metric | Value | Model |
|---|---|---|---|---|
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 41.7 | ResNet-101 (lpf3) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 46.1 | ResNet-101 (lpf3) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 35.7 | ResNet-101 (lpf3) |
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 41.5 | ResNet-50 (lpf5) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 45.3 | ResNet-50 (lpf5) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 35.2 | ResNet-50 (lpf5) |
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 41.1 | ResNet-101 (lpf2) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 45.2 | ResNet-101 (lpf2) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 35.1 | ResNet-101 (lpf2) |
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 41 | ResNet-101 (lpf5) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 45.8 | ResNet-101 (lpf5) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 34.8 | ResNet-101 (lpf5) |
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 40.3 | ResNet-50 (lpf2) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 45.1 | ResNet-50 (lpf2) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 33.4 | ResNet-50 (lpf2) |
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 40 | ResNet-50 (lpf3) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 44.7 | ResNet-50 (lpf3) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 34.3 | ResNet-50 (lpf3) |
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 38.7 | DenseNet121 (lpf5) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 42.7 | DenseNet121 (lpf5) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 32 | DenseNet121 (lpf5) |
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 38.3 | ResNet-34 (lpf2) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 42.8 | ResNet-34 (lpf2) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 32.4 | ResNet-34 (lpf2) |
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 38.3 | DenseNet-121 (lpf3) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 42.8 | DenseNet-121 (lpf3) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 32.3 | DenseNet-121 (lpf3) |
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 38.3 | ResNet-34 (lpf3) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 42.9 | ResNet-34 (lpf3) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 31.9 | ResNet-34 (lpf3) |
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 38.3 | DenseNet-121 (lpf2) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 43.1 | DenseNet-121 (lpf2) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 31.7 | DenseNet-121 (lpf2) |
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 37.2 | VGG-16 BN (lpf2) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 41.8 | VGG-16 BN (lpf2) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 31.3 | VGG-16 BN (lpf2) |
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 37.2 | ResNet-34 (lpf5) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 42.5 | ResNet-34 (lpf5) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 29.9 | ResNet-34 (lpf5) |
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 37 | VGG-16 BN (lpf5) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 41.7 | VGG-16 BN (lpf5) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 30.8 | VGG-16 BN (lpf5) |
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 36.9 | VGG-16 BN (lpf3) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 42.1 | VGG-16 BN (lpf3) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 30.6 | VGG-16 BN (lpf3) |
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 36 | MobileNetV2 (lpf3) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 40.3 | MobileNetV2 (lpf3) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 30.4 | MobileNetV2 (lpf3) |
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 35.8 | MobileNetV2 (lpf5) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 40.1 | MobileNetV2 (lpf5) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 29.1 | MobileNetV2 (lpf5) |
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 35.6 | ResNet-18 (lpf3) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 39.5 | ResNet-18 (lpf3) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 28.5 | ResNet-18 (lpf3) |
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 35.5 | MobileNetV2 (lpf2) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 39.2 | MobileNetV2 (lpf2) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 30.3 | MobileNetV2 (lpf2) |
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 35.5 | ResNet-18 (lpf2) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 40.1 | ResNet-18 (lpf2) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 28.7 | ResNet-18 (lpf2) |
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 35.1 | VGG-16 (lpf3) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 40 | VGG-16 (lpf3) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 28.2 | VGG-16 (lpf3) |
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 34.7 | ResNet-18 (lpf5) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 38.9 | ResNet-18 (lpf5) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 27.7 | ResNet-18 (lpf5) |
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 34.5 | VGG-16 (lpf5) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 39.4 | VGG-16 (lpf5) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 27.8 | VGG-16 (lpf5) |
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 33.5 | VGG-16 (lpf2) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 38.5 | VGG-16 (lpf2) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 26.7 | VGG-16 (lpf2) |
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 23.1 | AlexNet (lpf3) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 26.8 | AlexNet (lpf3) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 17.5 | AlexNet (lpf3) |
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 22.8 | AlexNet (lpf2) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 26.8 | AlexNet (lpf2) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 18.2 | AlexNet (lpf2) |
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 22.7 | AlexNet (lpf5) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 26.8 | AlexNet (lpf5) |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 18.4 | AlexNet (lpf5) |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 41.7 | ResNet-101 (lpf3) |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 46.1 | ResNet-101 (lpf3) |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 35.7 | ResNet-101 (lpf3) |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 41.5 | ResNet-50 (lpf5) |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 45.3 | ResNet-50 (lpf5) |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 35.2 | ResNet-50 (lpf5) |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 41.1 | ResNet-101 (lpf2) |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 45.2 | ResNet-101 (lpf2) |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 35.1 | ResNet-101 (lpf2) |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 41 | ResNet-101 (lpf5) |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 45.8 | ResNet-101 (lpf5) |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 34.8 | ResNet-101 (lpf5) |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 40.3 | ResNet-50 (lpf2) |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 45.1 | ResNet-50 (lpf2) |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 33.4 | ResNet-50 (lpf2) |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 40 | ResNet-50 (lpf3) |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 44.7 | ResNet-50 (lpf3) |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 34.3 | ResNet-50 (lpf3) |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 38.7 | DenseNet121 (lpf5) |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 42.7 | DenseNet121 (lpf5) |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 32 | DenseNet121 (lpf5) |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 38.3 | ResNet-34 (lpf2) |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 42.8 | ResNet-34 (lpf2) |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 32.4 | ResNet-34 (lpf2) |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 38.3 | DenseNet-121 (lpf3) |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 42.8 | DenseNet-121 (lpf3) |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 32.3 | DenseNet-121 (lpf3) |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 38.3 | ResNet-34 (lpf3) |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 42.9 | ResNet-34 (lpf3) |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 31.9 | ResNet-34 (lpf3) |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 38.3 | DenseNet-121 (lpf2) |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 43.1 | DenseNet-121 (lpf2) |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 31.7 | DenseNet-121 (lpf2) |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 37.2 | VGG-16 BN (lpf2) |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 41.8 | VGG-16 BN (lpf2) |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 31.3 | VGG-16 BN (lpf2) |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 37.2 | ResNet-34 (lpf5) |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 42.5 | ResNet-34 (lpf5) |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 29.9 | ResNet-34 (lpf5) |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 37 | VGG-16 BN (lpf5) |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 41.7 | VGG-16 BN (lpf5) |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 30.8 | VGG-16 BN (lpf5) |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 36.9 | VGG-16 BN (lpf3) |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 42.1 | VGG-16 BN (lpf3) |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 30.6 | VGG-16 BN (lpf3) |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 36 | MobileNetV2 (lpf3) |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 40.3 | MobileNetV2 (lpf3) |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 30.4 | MobileNetV2 (lpf3) |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 35.8 | MobileNetV2 (lpf5) |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 40.1 | MobileNetV2 (lpf5) |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 29.1 | MobileNetV2 (lpf5) |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 35.6 | ResNet-18 (lpf3) |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 39.5 | ResNet-18 (lpf3) |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 28.5 | ResNet-18 (lpf3) |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 35.5 | MobileNetV2 (lpf2) |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 39.2 | MobileNetV2 (lpf2) |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 30.3 | MobileNetV2 (lpf2) |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 35.5 | ResNet-18 (lpf2) |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 40.1 | ResNet-18 (lpf2) |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 28.7 | ResNet-18 (lpf2) |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 35.1 | VGG-16 (lpf3) |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 40 | VGG-16 (lpf3) |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 28.2 | VGG-16 (lpf3) |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 34.7 | ResNet-18 (lpf5) |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 38.9 | ResNet-18 (lpf5) |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 27.7 | ResNet-18 (lpf5) |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 34.5 | VGG-16 (lpf5) |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 39.4 | VGG-16 (lpf5) |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 27.8 | VGG-16 (lpf5) |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 33.5 | VGG-16 (lpf2) |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 38.5 | VGG-16 (lpf2) |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 26.7 | VGG-16 (lpf2) |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 23.1 | AlexNet (lpf3) |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 26.8 | AlexNet (lpf3) |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 17.5 | AlexNet (lpf3) |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 22.8 | AlexNet (lpf2) |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 26.8 | AlexNet (lpf2) |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 18.2 | AlexNet (lpf2) |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 22.7 | AlexNet (lpf5) |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 26.8 | AlexNet (lpf5) |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 18.4 | AlexNet (lpf5) |