Log-Polar Mapping Overview

• Log-polar mapping is a nonlinear coordinate transformation that redefines points using the logarithm of radial distance and angular coordinates, offering foveated sampling similar to the primate retina.
• It mathematically converts Cartesian coordinates to a log-polar domain, enabling scale and rotation equivariance through cyclic angular shifts and logarithmic radial adjustments.
• The transformation is widely applied in computational vision, deep learning, and tracking, though it poses challenges such as singularities and nonuniform sampling density.

Log-polar mapping is a nonlinear coordinate transformation that replaces the standard Cartesian (x, y) representation of points in the plane with a system based on the logarithm of the radial distance and an angular coordinate. Formally, given a reference center $(x_0, y_0)$, a point is mapped as $$(x, y) \mapsto (\rho = \ln \sqrt{(x - x_0)^2 + (y - y_0)^2}, \theta = \operatorname{atan2}(y - y_0, x - x_0))$$. This mapping produces a domain with exponentially expanding sampling in radius, yielding high density near the origin and coarse sampling in the periphery—a property directly inspired by the primate foveal system. Log-polar mapping has been applied extensively in computational vision, deep learning architectures, radar/sonar tracking, local descriptor learning, and fast integral transforms.

1. Mathematical Foundations of Log-Polar Mapping

The log-polar transform proceeds in two principal steps. First, Cartesian coordinates are mapped to standard polar form: $$r = \sqrt{(x - x_0)^2 + (y - y_0)^2}, \quad \theta = \operatorname{atan2}(y - y_0, x - x_0)$$ Then, the radial variable is mapped logarithmically: $\rho = \ln(r + r_0)$ where $r_0 \geq 0$ is an optional offset to avoid singularity at $r=0$ (Kiritani et al., 2020, Remmelzwaal et al., 2019). The inverse transform reconstructs Cartesian coordinates: $$x = x_0 + (e^\rho - r_0) \cos \theta,\quad y = y_0 + (e^\rho - r_0) \sin \theta$$ This formulation underlies all analytic log-polar mappings, regardless of application domain.

Discretization proceeds by uniformly sampling $\rho$ and $\theta$ to construct a grid, resulting in $N_\rho \times N_\theta$ bins that form the domain of log-polar images or patches (Remmelzwaal et al., 2019, Bhowmik et al., 2010, Kiritani et al., 2020). In feature extraction contexts, quantization in radial and angular directions produces concentric rings and angular sectors, forming a support structure that oversamples at the center and under-samples at large radii (Ebel et al., 2019).

Bilinear interpolation is typically used during resampling to assign values to non-integer pixel locations in the original grid (Remmelzwaal et al., 2019, Ebel et al., 2019, Kiritani et al., 2020). Nearest-neighbor interpolation may also be employed for computational simplicity in high-throughput settings (Bhowmik et al., 2010).

2. Biological and Computational Motivation

Log-polar sampling is strongly motivated by the structure of the vertebrate retina, where the density of photoreceptors is highest at the fovea and drops off logarithmically with eccentricity (Remmelzwaal et al., 2019, Kiritani et al., 2020). This arrangement enables high spatial acuity near fixation and efficient, coarse monitoring of the visual periphery. Log-polar systems mimic these properties, offering high central spatial resolution and rapid falloff, which are attractive for attentive systems and resource-constrained perception.

These properties are reflected in artificial systems. For instance, the Recurrent Attention Model with Log-Polar Mapping (RAM-LPM) exploits a log-polar glimpse sensor whose field-of-view is dynamically controlled by an attention mechanism, resulting in both biological plausibility and computational efficiency (Kiritani et al., 2020). Similarly, multi-resolution foveated "log-polar-like" sensors in robotic gaze control use nested, downsampled regions of interest to approximate foveated sampling with minimal computational burden (Göransson et al., 2023).

3. Implementation in Machine Vision and Deep Learning

Log-polar mapping is a core primitive in multiple deep learning and machine vision architectures:

• Preprocessing: Direct application of log-polar transformation as a preprocessor yields substantial rotation and scale robustness. CNNs trained on such warped images maintain classification accuracy under extensive geometric transformations and support data compression by factors of 5× with minimal loss (Remmelzwaal et al., 2019, Bhowmik et al., 2010).
• Deep descriptor learning: Incorporation of log-polar patch extraction layers in deep local descriptor pipelines enables robust matching across large scale and orientation changes, outperforming Cartesian baselines. This is implemented as a differentiable warp based on keypoint position, scale, and orientation, followed by bilinear interpolation and batchwise sampling (Ebel et al., 2019).
• Convolutional kernel design: Log-Polar Space Convolution (LPSC) replaces square spatial kernels with elliptically-shaped kernels whose bins tile log-polar rings and angular sectors about a center. This allows for exponential growth of receptive field with only a linear increase in parameter count, enabling architectures such as AlexNet, VGG-19, and ResNet-18 to use vastly enlarged kernels without parameter penalty (Su et al., 2021).

A table summarizes core implementation strategies and architectural uses:

Application	Log-Polar Usage	Key Reference
Preprocessing for CNNs	Fixed remap and sampling	(Remmelzwaal et al., 2019, Bhowmik et al., 2010)
Attention/Motion modules	Dynamic log-polar glimpses	(Kiritani et al., 2020, Göransson et al., 2023)
Descriptor learning	Learnable patch extraction	(Ebel et al., 2019)
Large-kernel convolutions	LPSC binning and pooling	(Su et al., 2021)

4. Equivariance, Robustness, and Theoretical Properties

Log-polar mapping imparts an approximate equivariance to rotation and scale, a property central to its utility. Under rotation about the center, $\Delta\phi$, the angular coordinate shifts: $\theta \to \theta + \Delta\phi$. In the log-polar grid, this is a cyclic shift along the angle axis, which convolutional filters handle naturally. Uniform scaling, $r \to s r$, translates to $\rho \to \rho + \ln s$, a vertical shift in the log-radius axis (Remmelzwaal et al., 2019, Bhowmik et al., 2010, Ebel et al., 2019, Kiritani et al., 2020).

CNNs' inherent translation equivariance thus extends to scale and rotation invariance in the log-polar domain, reducing the necessity for extensive data augmentation or architectural modifications. Experimental evidence confirms that downstream networks maintain classification accuracy or descriptor quality under wide variations in rotation and scale after log-polar preprocessing; error rates drop by 8–10 percentage points in holistic face recognition and maintain >90% MNIST accuracy with 60% scaling and arbitrary rotation (Remmelzwaal et al., 2019, Bhowmik et al., 2010).

In adversarial robustness, models using log-polar glimpses exhibit enhanced resistance to $\ell^p$ attacks relative to standard CNNs, attributable to the combination of foveated sampling and the transformation properties above (Kiritani et al., 2020).

5. Applications Beyond Image Classification

Log-polar mapping supports advanced applications in tracking, signal processing, and attention-driven control:

• Bearings-only tracking: State estimation algorithms, such as the Unscented Kalman Filter (UKF), benefit from log-polar coordinates by maintaining positivity of range (since the log-normal distribution does not permit negative radii) and enabling closed-form updates of posterior mean, covariance, and even third- and fourth-order moments under specific motion models (Xiourouppa et al., 2024).
• Integral transforms: The hyperbolic Radon transform, critical in seismic data analysis, is recast from summation over hyperbolic curves to efficient 2D convolutions using log-polar coordinates, allowing for $\mathcal O(N^2 \log N)$ implementations using FFTs and GPU infrastructure (Nikitin et al., 2016).
• Robotic and reinforcement learning agents: Multi-resolution log-polar or "log-polar-like" sensors efficiently compress input space for real-time control, yielding 5× reduction in raw visual data without sacrificing policy effectiveness in Atari gaming and robot gaze control (Göransson et al., 2023).

6. Limitations and Practical Considerations

Notable limitations include:

• Singularity at the origin: The logarithm diverges at $r=0$, necessitating either an additive offset ($r_0$) or exclusion of the singular point (Kiritani et al., 2020, Remmelzwaal et al., 2019, Bhowmik et al., 2010).
• Nonuniform sampling density: Physical space bin widths grow with radius, potentially distorting global geometry and undersampling elongated features in the periphery (Bhowmik et al., 2010).
• Interpolation artifacts: Resampling via nearest-neighbor or bilinear interpolation can introduce aliasing or smooth out details, especially near bin boundaries for aggressive compression (Bhowmik et al., 2010, Remmelzwaal et al., 2019).
• FOV design trade-offs: Selection of center, inner/outer radius bounds, and angular/radial resolutions involves a trade-off between capturing global structure and central acuity; mis-specification can degrade performance on large or peripheral objects (Kiritani et al., 2020).
• Algorithmic complexity in attention/RL: For models requiring hard attention and REINFORCE, such as RAM-LPM, training may be sample-inefficient and hyperparameter-sensitive (Kiritani et al., 2020).

7. Extensions and Future Directions

Emerging directions include:

• End-to-end differentiable log-polar layers: Incorporating log-polar warping as parameterized layers with gradients allows networks to optimize over FOV parameters or to backpropagate through the mapping itself (Kiritani et al., 2020).
• Hybrid attention and multiscale designs: Integrations of soft and hard attention mechanisms or stacking of multiple log-polar fields-of-view can potentially alleviate combinatorial exploration burdens in RL and sequential vision architectures (Kiritani et al., 2020).
• Support region learning: Dynamic determination of log-polar parameters (e.g., learnable $r_0$, radii steps, or center points) may further improve robustness to scene structure and support scale-aware descriptors (Ebel et al., 2019).
• Signal processing and tracking: The availability of higher-order analytic moments in log-polar coordinates facilitates non-Gaussianity monitoring and adaptive filtering beyond the Kalman regime, as demonstrated by closed-form third/fourth central moments in bearings-only tracking (Xiourouppa et al., 2024).

Log-polar mapping thus constitutes a mathematically principled, biologically grounded, and computationally effective strategy for achieving rotation, scale, and peripheral robustness alongside fixed computational cost in perceptual and control systems. Its integration into modern deep networks and signal processing pipelines continues to propagate as data growth, high-dimensional signals, and real-time constraints increasingly demand efficient, equivariant, and foveated representations.