Polar Transformer Networks and warped convolutions use a simple geometric fact. If you resample an image on a log-polar grid (rings of radius growing exponentially around a center point, crossed by evenly spaced angular spokes), then rotating the image becomes a shift along the angle axis, and scaling it becomes a shift along the log-radius axis. An ordinary convolution on that grid therefore responds the same way to rotated and rescaled inputs. That symmetry has a cost these methods do not discuss. One filter has to serve every ring, and the rings are not statistically alike. Near the center, each log-polar cell covers less than a pixel, so the sampler interpolates and blurs. Far out, each cell covers many pixels, so the sampler aliases unless it smooths first. A filter shared across rings is being asked to detect the same edge in an over-smoothed signal and in an aliased one.
We propose to keep sharing exact where the geometry is exact and to relax it where the geometry is not. The layer uses one filter shared around each ring, so rotation equivariance is kept exactly. Filters are allowed to differ from ring to ring, through a penalized deviation from a common filter. A single knob moves the layer between a Polar Transformer Network (full sharing) and a radially locally connected layer, which is a layer with separate weights at each position but only along the radial axis. The number of free filters grows with the number of rings, roughly 32, rather than with the full grid area. The scientific question is when it pays to give up scale equivariance in exchange for matching the grid's ring-dependent statistics. We also ask whether those gains come from sampling artifacts that better anti-aliasing would remove, or from real differences between the center and the periphery of the image.
Core Claim
On a log-polar grid, untying weights along the radius and sharing them along the angle beats fully shared log-polar convolution when test scales lie inside the training range. The gain is largest when the warp is not anti-aliased. At extreme unseen scales, the shared model should keep its advantage, and the knob should trace an explicit trade-off between the two regimes rather than a free lunch.
Research Program
The layer acts on a feature map indexed by ring $i \in \{1,\dots,N_r\}$ and angle $j \in \{1,\dots,N_\theta\}$. The filter applied at ring $i$ is $W_i = W_0 + \Delta_i$. Here $W_0$ is a shared $k\times k$ filter bank and $\Delta_i$ is a ring-specific correction. The correction is built from $M$ learned basis filters $B_m$ as $\Delta_i = \sum_{m=1}^{M} a_{i,m} B_m$, where $a_{i,m}$ are per-ring mixing coefficients. Training adds the penalty
$$ \mathcal{R} = \lambda \sum_{i} \|\Delta_{i+1}-\Delta_i\|^2 + \mu \sum_i \|\Delta_i\|^2, $$
where $\lambda$ makes the corrections vary smoothly from ring to ring and $\mu$ pulls them toward zero. Large $\mu$ recovers the shared convolution of Polar Transformer Networks. Taking $\mu=\lambda=0$ with $M$ at full rank gives a radially locally connected layer.
Convolution along $\theta$ is cyclic, using the wrap-around padding from the Polar Transformer Network paper. Because the weights depend only on $i$, rotating the input by any multiple of $2\pi/N_\theta$ cyclically shifts the output by exactly the same amount. The center of the warp comes from that paper's differentiable origin predictor, a small network whose output heatmap's centroid becomes the warp center.
Each prior line of work plays a different role here. Warped Convolutions and Polar Transformer Networks supply the warp and the equivariance argument. Both apply globally shared convolutions after the warp and never question whether the warped grid is stationary, that is, whether its local statistics are the same at every position. That assumption is what we test.
The neuroscience work supplies the motivation. Explaining V1 Properties with a Biologically Constrained Deep Learning Architecture found that a log-polar front end and untied local receptive fields were both needed to match primary visual cortex (V1). It untied weights over the whole grid, used a fixed center, and lost ImageNet accuracy (61% to 55%). Our layer unties only along one axis, learns the center, and is tested on geometric tasks. The emergent center-to-periphery gradients that End-to-end topographic networks found in fully untied Cartesian networks suggest that ring-dependent processing is something networks want, but those networks pay the full area-scaling parameter cost.
Other work constrains the design. Locally connected networks as ventral stream models reports a large accuracy loss from full untying on ImageNet (44.9% versus 69.9% for ResNet-18). This is why we start corrections at zero and penalize them rather than untie freely. Learning in the Machine: To Share or Not to Share? shows that untied weights trained with augmentation drift toward shared values. That suggests the learned corrections $\Delta_i$ will shrink wherever sharing is actually correct, which makes their size a useful diagnostic.
Log-Polar Space Convolution Layers is the nearest name match, but it does something different. It gives the inside of a kernel a log-polar layout and slides that one kernel over a Cartesian image. No weights vary with position.
For analysis, we will measure the size of each correction against ring index. We will then check whether that profile follows the sampler's pixels-per-cell curve, which would mean the model is compensating for sampling, or follows content statistics, which would mean real differences between center and periphery.
Baselines
The naive reference point is a standard Cartesian CNN trained with rotation and scale augmentation. Polar Transformer Networks reports it at 11.73% error on SIM2MNIST, against 12.35% for a Spatial Transformer Network. We will match its parameter count to our model and give it the same augmentation and tuning budget.
The strongest existing method is the Polar Transformer Network itself. It reports 0.89% error on Rotated MNIST and 5.03% on SIM2MNIST. We will re-implement it inside our own code, so that the only difference from our model is the regularizer. This uses identical origin predictor, grid resolution, depth, optimizer, and a matched hyperparameter search of about 30 trials per method. Beating it means lower error on in-range test scales, averaged over 5 seeds, with non-overlapping confidence intervals. We also require that the gain survives the anti-aliased variant described next.
On the Google Earth vehicle pose dataset, the reference is the warped CNN of Warped Convolutions at 8.29° rotation error. A standard CNN with a dense head reaches 22.54° on the same task.
Three ablations isolate the key ingredient. The first is full untying ($\mu=\lambda=0$), which tests whether the regularizer matters at all. The second unties along the angle instead of the radius, which breaks rotation equivariance and should hurt; this checks that the choice of axis matters. The third trains each method on a warp with radius-dependent pre-smoothing in the style of Scale invariance without scale selection. That paper uses spatially varying smoothing so that the warped signal becomes closer to stationary. This third ablation separates "the model fixes the sampler" from "the model exploits content."
Why not just…?
Why not just append a log-radius coordinate channel to the shared convolution? This is the most serious threat. A coordinate input lets a shared network become position-aware at almost no cost. It enters additively, though, and mainly at the input. Our corrections change each layer's filters multiplicatively. We will run the coordinate-channel variant as a full baseline under the same budget. If it matches ring-untying, the honest conclusion is that knowing your position is enough and separate per-ring weights are unnecessary. We would report that.
Why not just anti-alias the warp properly and keep weights shared? If the non-stationarity comes entirely from the sampler, a better prefilter should remove it. The third ablation tests exactly this. If untying gains disappear on the pre-smoothed warp, the contribution shrinks to a diagnosis of a sampling flaw in Polar Transformer Networks plus its fix. That is still useful, but it is a different paper. A gain that survives pre-smoothing would point to genuine center-versus-periphery content differences.
Why not just train a Cartesian CNN with heavy augmentation? The evidence in the file already argues against it. The 11.73% versus 5.03% gap on SIM2MNIST, and the 22.54° versus 8.29° gap on vehicle pose, show that augmentation does not substitute for the warp. We keep it as the naive floor.
Why not untie over the whole grid, as in the V1 models? That approach costs a separate filter at every grid position. It also gives up rotation equivariance. The ImageNet losses in both V1 papers indicate it overfits. The full-untying ablation measures this cost directly.
Proposed Experiments
We will train on Rotated MNIST, SIM2MNIST, and the Google Earth vehicle pose dataset. We will also build a controlled scale-extrapolation split of MNIST: train on scale factors in [0.7, 1.4], test on that range and separately on [0.4, 0.7) and (1.4, 2.5]. Metrics are classification error, or rotation error in degrees for pose, plus error as a function of test scale.
The decisive check is a family of curves of error versus test scale, swept over $\mu$ and repeated with and without pre-smoothing. The hypothesis is confirmed if moderate $\mu$ beats the Polar Transformer Network in range without losing more than its own advantage out of range. It is refuted if the in-range gains vanish once the warp is pre-smoothed or once the coordinate-channel baseline is added. A secondary, exploratory check is to swap our layer into the front end of the V1 model and score it on Brain-Score V1, against that paper's 0.629.
References
- Polar Transformer Networks
- Warped Convolutions: Efficient Invariance to Spatial Transformations
- Explaining V1 Properties with a Biologically Constrained Deep Learning Architecture
- End-to-end topographic networks as models of cortical map formation and human visual behaviour: moving beyond convolutions
- Locally connected networks as ventral stream models
- Learning in the machine: To share or not to share?
- Log-Polar Space Convolution Layers
- Scale invariance without scale selection