Image networks change their predictions when the input moves by one pixel, or even a fraction of one. The best current fixes remove this sensitivity almost entirely, but only under an artificial notion of "shift": the image wraps around like a torus, so pixels that leave on the right re-enter on the left. Those fixes also do their filtering with whole-image Fourier transforms, which quietly assume the image is periodic. Real camera motion is a crop-shift instead: new content enters at one border and old content leaves at the other. Under a crop-shift, a global Fourier filter treats the mismatch between opposite image edges as a sharp jump. It then spreads ringing from that jump across the entire feature map, so a change at the border disturbs predictions far from it.
We propose to do all anti-aliasing with compactly supported filters, meaning filters that are exactly zero beyond a few pixels. Here anti-aliasing means removing frequencies too high to survive a reduction in resolution. The filters we use are B-splines, smooth piecewise polynomials with finite width. They are applied at three points: before every downsampling step, around the nonlinear activations, and in the interpolation used for sub-pixel shifts. The claim is that this makes the residual error from a crop-shift local. Pixels farther from the image border than the network's receptive field should behave equivariantly up to a small, controllable spline error. That error can be traded against filter width by choosing the spline order. Global methods cannot offer this. Dense prediction on large images, such as street-scene segmentation, has a real "interior" that this guarantee can protect.
Core Claim
Fix a crop-shift by $t$ pixels, possibly fractional. Write $\Delta(p)$ for the change in a dense network's output at pixel $p$, after undoing the shift. If every operation in the network has finite support and the total receptive-field radius is $R$, then pixels with border distance $d(p) > R$ see only approximation error. That error comes from the spline interpolation and the imperfect low-pass filtering, not from the boundary. For a degree-$n$ B-spline, the frequency response is a sinc raised to the power $n+1$. Energy leaking from above the downsampled grid's Nyquist limit (half the sampling rate) therefore falls off roughly as $|\omega|^{-(n+1)}$, where $\omega$ is frequency, while the filter's width grows linearly in $n$. Together this gives the prediction
$$\mathbb{E}\,[\Delta(p) \mid d(p) > R(n)] \le \varepsilon(n), \qquad \varepsilon(n) \text{ decreasing in } n,\; R(n) \text{ increasing in } n.$$
This is the space–frequency trade-off of compact support, stated as something one can measure. Global FFT-based filters are, in effect, the $n \to \infty$ limit. In that limit $\varepsilon$ is nearly zero under circular shifts, but $R$ becomes the whole image. The proposal is to find the spline order at which crop-shift consistency in the interior is best, and to show that it beats both naive networks and FFT-based alias-free networks there, at a fraction of their cost.
Position Relative to Prior Work
The motivation comes from three findings. First, small translations flip labels and make detection boxes jitter (Azulay; Manfredi et al.). Second, When Semantic Segmentation Meets Frequency Aliasing shows that segmentation errors at object boundaries track an Aliasing Score. That score is computed from an Equivalent Sampling Rate, an estimate of the effective sampling rate of each layer. Third, gradient descent does not learn the missing low-pass filters on its own: small kernels lack the spatial support to implement them (Vasconcelos et al.), and networks memorize aliased patterns across channels rather than filtering them out (Ribeiro et al.).
The strongest competitors all share one assumption. Alias-Free Convnets replaces activations with polynomials, whose bandwidth expansion can be tracked exactly, and filters with FFTs. Alias-Free ViT truncates frequencies with FFTs in the patch embedding. Chen et al.'s De-Aliasing Filter zeroes frequencies above the cutoff with whole-map FFTs. Learnable Polyphase Sampling instead learns which sub-grid to keep when downsampling. All four certify consistency only under circular shifts, and the file reports that consistency degrades under crop-shifts. The two alias-free networks also cost about 5–7× more to train.
We keep their goal, removing aliasing at both downsampling and activations, but replace global transforms with finite filters. We also measure the quantity they leave uncertified: crop-shift consistency as a function of distance from the border.
Making Convolutional Networks Shift-Invariant Again already uses compact blur kernels. Its binomial kernels are essentially discrete B-splines, so the overlap with our method is real. The difference is that BlurPool ignores activation aliasing and sub-pixel shifts, and it was never analyzed for where its errors fall spatially.
Two papers supply methods we reuse. Stable parameterization of continuous and piecewise-linear functions shows that local hat and box-spline bases are well conditioned, with Riesz condition number $\sqrt{d+2}$ on uniform grids. We rely on this when we make spline filter coefficients learnable. The bandwidth-tracking logic of the Alias-Free Convnet tells us where to place filters around activations.
Two lines of work constrain the design. On Translation Invariance in CNNs and Mind the Pad show that zero-padding leaks absolute position into the network, so we use mirror padding throughout and treat padding as a controlled variable. From CNNs to Shift-Invariant Twin Wavelet Models and the wavelet pooling of The Importance of Anti-Aliasing in Tiny Object Detection show that compact wavelet filters can replace pooling with no loss in accuracy. However, the first touches only the network's first layer, and the second discards high-frequency subbands without measuring shift consistency.
Research Program
We will build a drop-in "spline-alias-free" variant of a ResNet-50 backbone inside the UPerNet segmentation model, the setting used by Chen et al. The variant changes three things.
At downsampling, every strided operation becomes a dense evaluation, then a separable degree-$n$ B-spline prefilter, then subsampling. This is BlurPool, with the spline order exposed as a parameter that sets the trade-off.
At each activation in the stages that feed a downsampling step, the feature map is upsampled 2× by spline interpolation, passed through ReLU, then spline-prefiltered and downsampled. This is the Alias-Free Convnet's recipe with FFTs replaced by finite filters. Limiting it to stages before downsampling keeps the cost bounded, and we will measure that overhead directly.
For sub-pixel evaluation, fractional crop-shifts are produced by cubic B-spline interpolation of the raw image. This is applied identically to the test input and to the reference used to undo the shift, so the measurement does not favor any method.
We will first verify the interior-locality claim analytically and with synthetic inputs, by perturbing only a border strip and mapping where outputs change. We will then sweep spline orders 1 to 7 and record two curves: interior consistency $\varepsilon(n)$ and the radius $R(n)$ at which consistency stops depending on border distance. A learnable variant will initialize spline coefficients at the fixed B-spline values and fine-tune them under a penalty on the Aliasing Score. This tests whether learning recovers detail that fixed blurs remove, which Chen et al. address with their frequency-mixing module.
Baselines
The naive reference point is the unmodified UPerNet-R50 with standard strided convolutions, max-pooling, and zero-padding, trained with the same schedule. Chen et al. report improvements relative to this model rather than its absolute mIoU in the file, so we will reproduce it ourselves.
The strongest existing method for segmentation under aliasing is Chen et al.'s De-Aliasing Filter plus Frequency Mixing on UPerNet-R50. They report +1.6% mIoU on Cityscapes and +1.8% on PASCAL VOC over that baseline. Beating it means matching its mIoU within noise while showing clearly higher crop-shift consistency at border distances beyond $R(n)$. The strongest method for shift consistency in dense prediction is Learnable Polyphase Sampling with its upsampling counterpart. It reports 100% circular-shift consistency (mASCC) on PASCAL VOC. We found no reported result on Lacuna for its crop-shift or fractional-shift consistency on Cityscapes or VOC, and we will measure both.
We also include BlurPool with mirror padding, since it is the compact-support method closest to ours, and an Alias-Free-Convnet-style variant. The latter uses FFT activation filtering with mirror padding in place of circular convolution, isolating whether global filtering is the problem once the periodic boundary is removed.
The key ablation keeps our architecture identical but swaps every spline filter for an FFT truncation at the same cutoff. It isolates compact support as the ingredient. A second ablation removes the activation path, isolating the gain from treating activation aliasing over BlurPool alone.
All methods share the training data, crop size, schedule, augmentation, and mirror padding, except where padding is the variable under study. Each gets the same hyperparameter search budget, and we report wall-clock training and inference time. As a sanity check on classification, the Alias-Free Convnet reports 81.04% ImageNet top-1 with 100% consistency under circular integer and fractional shifts. We will not claim to beat that circular figure. We only check whether our backbone gives up accuracy for its locality.
Why not just…?
Why not just train with random crop-shift augmentation? Ribeiro et al. show that networks reduce training loss under aliasing by memorizing aliased patterns across channels rather than removing them. That suggests augmentation teaches the training distribution of shifts without delivering consistency elsewhere. This remains an empirical question, not a settled one. We will train the naive baseline and our model with heavy integer and fractional shift augmentation and compare consistency on held-out shift magnitudes. If augmentation alone closes the gap in the interior, the architectural claim fails.
Why not just use BlurPool with mirror padding, which is already compact? This is the most serious threat, because our downsampling filters are BlurPool-like. The Alias-Free Convnet's argument is that ReLU creates new high frequencies after the blur, so BlurPool should leave residual aliasing that grows under fractional shifts. The ablation that removes our activation path is exactly BlurPool with spline kernels. If it matches the full model on fractional crop-shift consistency, the contribution shrinks to the analysis and the order sweep, and we will report it as such.
Why not just run the FFT-based alias-free methods with reflect padding? Reflect padding removes the edge mismatch that drives ringing, so it may rescue global methods. However, an FFT low-pass filter still has sinc-like spatial tails, so the influence of the border decays slowly rather than stopping at a finite radius. The FFT-swap ablation and the reflect-padded FFT baseline test this directly by plotting consistency against border distance. The global methods also keep their reported 5–7× training cost, which our separable finite filters avoid.
Proposed Experiments
The main benchmark is Cityscapes semantic segmentation. Its 1024×2048 images leave a large interior relative to ResNet-50's receptive field. PASCAL VOC serves as a second dataset that matches the reported numbers of Chen et al. and Learnable Polyphase Sampling.
Metrics are mIoU and boundary IoU for accuracy, plus the Aliasing Score per layer. For consistency we use the fraction of pixels whose predicted label is unchanged after integer and fractional crop-shifts of 0.25 to 8 pixels, binned by distance to the image border. For cost we report training and inference time.
Three outcomes would decide the question. The first is whether consistency in our model becomes flat beyond $R(n)$, while FFT variants stay border-dependent even with reflect padding. The second is whether interior consistency improves with spline order up to an optimum, as the trade-off predicts. The third is whether all of this holds at Chen et al.'s mIoU level with less than 2× training cost.
As a stress test, we repeat the evaluation under Gaussian noise. Chen et al. found that noise raises early-layer aliasing, so this is where low-pass filtering should matter most. A failure on the first check, meaning interior consistency still depends on border distance, would refute the locality claim outright.
References
- Why do deep convolutional networks generalize so poorly to small image transformations?
- Shift Equivariance in Object Detection
- When Semantic Segmentation Meets Frequency Aliasing
- Impact of Aliasing on Generalization in Deep Convolutional Networks
- How Convolutional Neural Networks Deal with Aliasing
- Alias-Free Convnets: Fractional Shift Invariance via Polynomial Activations
- Alias-Free ViT: Fractional Shift Invariance via Linear Attention
- Learnable Polyphase Sampling for Shift Invariant and Equivariant Convolutional Networks
- Making Convolutional Networks Shift-Invariant Again
- Stable parameterization of continuous and piecewise-linear functions
- On Translation Invariance in CNNs: Convolutional Layers Can Exploit Absolute Spatial Location
- Mind the Pad - CNNs Can Develop Blind Spots
- From CNNs to Shift-Invariant Twin Wavelet Models
- The Importance of Anti-Aliasing in Tiny Object Detection