Region-Aware Deformable Convolution (RAD-Conv)

• RAD-Conv is a convolution operator that aggregates features over learned, continuous rectangular regions rather than fixed grid point samples.
• It decouples receptive-field geometry from kernel size, enabling adaptive aspect ratios and flexible integration even with 1×1 kernels.
• The method employs exact continuous integration over bilinearly interpolated features, providing smooth and differentiable gradient propagation.

Region-Aware Deformable Convolution (RAD-Conv) is a convolutional operator introduced in “Region-Aware Deformable Convolutions” (Maleki et al., 18 Sep 2025). It replaces the deformable-convolution primitive of sampling a single offset point per kernel element with a boundary-parameterized rectangular region, defined by four learned offsets and aggregated by exact continuous integration over a bilinearly interpolated feature field. The operator is presented as a middle ground between rigid convolution, deformable convolution, and attention: it preserves convolutional weight sharing and structure, yet decouples receptive-field size and aspect ratio from kernel size, so that even a $1\times1$ kernel can aggregate over a large content-adaptive neighborhood. The manuscript is primarily conceptual and mathematical; it does not report benchmark results, datasets, or ablation studies (Maleki et al., 18 Sep 2025).

1. From fixed grids and point offsets to region integration

RAD-Conv is formulated against two prior baselines. In standard convolution, an output at location $\mathbf{p}_0$ is computed on a fixed stencil,

$$\mathbf{y}(\mathbf{p}_0) = \sum_{k=1}^{K} \mathbf{w}_k \cdot \mathbf{x}(\mathbf{p}_0 + \mathbf{p}_k),$$

where the kernel grid $\mathcal R=\{\mathbf p_k\}_{k=1}^K$ determines the receptive-field geometry. This yields translation equivariance and efficiency, but it ties the receptive field to the kernel structure. The $1\times1$ case is the limiting example,

$$\mathbf{y}(\mathbf{p}_0) = \mathbf{w} \cdot \mathbf{x}(\mathbf{p}_0),$$

which has no spatial modeling at all. Longer-range context must therefore be obtained by stacking layers or enlarging the kernel (Maleki et al., 18 Sep 2025).

Deformable convolution relaxes the fixed-grid assumption by learning offsets. In DCNv1,

$$\mathbf{y}(\mathbf{p}_0) = \sum_{k=1}^{K} \mathbf{w}_k \cdot \mathbf{x}(\mathbf{p}_0 + \mathbf{p}_k + \Delta \mathbf{p}_k),$$

with $\Delta \mathbf p_k\in\mathbb R^2$ predicted from the input. DCNv2 adds modulation, and DCNv3 further uses grouped channels and softmax-normalized modulation. Yet the paper’s critique is geometric rather than purely functional: despite continuous offsets, each sample remains a point sample reconstructed from exactly four neighboring lattice points through bilinear interpolation. The effective geometry per sample is therefore still a quadrilateral tied to the interpolation stencil. In this view, DCN changes where it samples, but does not independently control the width and height of the contributing region.

RAD-Conv is designed to remove that restriction. Its central change is to replace the primitive “sample one deformed point” with “integrate over one deformed region.” This recasts adaptive convolution from offsetting discrete sampling locations to predicting continuous, axis-aligned support regions whose extent can vary with image content.

2. Operator definition and mathematical structure

In RAD-Conv, each kernel element predicts four positive boundary offsets—top, bottom, left, and right—that define an axis-aligned rectangle around the nominal anchor location. The full grouped operator is written as

$$\mathbf{y}(\mathbf{p}_0) = \sum_{g=1}^{G} \sum_{k=1}^{K} \mathbf{w}_g \cdot \mathbf{m}_{gk} \cdot \mathbf{x}_{[t_{gk},b_{gk}][l_{gk},r_{gk}]},$$

where $\mathbf{x}_{[t_{gk},b_{gk}][l_{gk},r_{gk}]}$ denotes the aggregated feature over the rectangular region associated with group $\mathbf{p}_0$0 and kernel element $\mathbf{p}_0$1 (Maleki et al., 18 Sep 2025).

The region boundaries are parameterized relative to the kernel anchor $\mathbf{p}_0$2. The manuscript’s typesetting is corrupted at this point, but its intended meaning is stated explicitly: as in FCOS-style boundary regression, four positive distances from a reference point define the top, bottom, left, and right sides of a valid rectangle. Positivity is enforced so that top lies above bottom and left lies left of right. The resulting parameterization is therefore a boundary-distance parameterization rather than an unconstrained corner-coordinate regression.

Once the rectangle has been defined, RAD-Conv does not extract a sparse set of samples inside it. Instead, it computes the average feature over the continuous rectangular region,

$\mathbf{p}_0$3

where $\mathbf{p}_0$4 is the continuous feature map induced from the discrete feature lattice. The paper states that this integral is computed analytically following Precise RoI Pooling, that is, by exact continuous integration over the bilinearly interpolated feature field rather than by quantized bin pooling. In consequence, RAD-Conv is closer to continuous region pooling embedded inside convolution than to ordinary point sampling.

Operationally, each kernel element and channel group executes the same sequence: it takes the nominal kernel location, predicts four boundary offsets, forms a rectangle, computes the average feature inside that rectangle, scales the result with a dynamic modulation weight, applies the learned convolution weight, and contributes the result to the output sum. The aggregation primitive is thus continuous, rectangular, and content-adaptive.

3. Geometric interpretation and receptive-field behavior

The geometric significance of RAD-Conv lies in decoupling receptive-field geometry from kernel geometry. In standard convolution, the kernel determines both the number of sampled locations and the spatial support pattern. In deformable convolution, the kernel points can move, but the operation still consists conceptually of point samples arranged according to the kernel. In RAD-Conv, by contrast, the kernel supplies only the number of region slots $\mathbf{p}_0$5; each slot independently determines what region it observes (Maleki et al., 18 Sep 2025).

This independence is expressed in four controllable quantities per kernel element: anchor position, width $\mathbf{p}_0$6, height $\mathbf{p}_0$7, and aspect ratio through the relative magnitudes of those extents. The paper emphasizes that such independent width/height control matters for anisotropic structures, including thin elongated edges, horizontally spread context, vertically extended patterns, and compact local details. A point-sampling offset cannot directly specify “use a tall narrow region here” and “a short wide region there” elsewhere, whereas RAD-Conv can.

A central special case is the $\mathbf{p}_0$8 kernel. RAD-Conv still retains spatial modeling power: $\mathbf{p}_0$9 The manuscript highlights this as one of its main claims. Standard $$\mathbf{y}(\mathbf{p}_0) = \sum_{k=1}^{K} \mathbf{w}_k \cdot \mathbf{x}(\mathbf{p}_0 + \mathbf{p}_k),$$0 convolution has no spatial support, and deformable $$\mathbf{y}(\mathbf{p}_0) = \sum_{k=1}^{K} \mathbf{w}_k \cdot \mathbf{x}(\mathbf{p}_0 + \mathbf{p}_k),$$1 convolution can only sample a single shifted point, effectively reconstructed from four neighboring lattice points. RAD-Conv, however, can aggregate over a large, content-adaptive region even with a $$\mathbf{y}(\mathbf{p}_0) = \sum_{k=1}^{K} \mathbf{w}_k \cdot \mathbf{x}(\mathbf{p}_0 + \mathbf{p}_k),$$2 kernel. This is the mechanism by which receptive-field size and aspect ratio are detached from kernel size.

The paper also positions RAD-Conv relative to attention. Global attention is described by

$$\mathbf{y}(\mathbf{p}_0) = \sum_{k=1}^{K} \mathbf{w}_k \cdot \mathbf{x}(\mathbf{p}_0 + \mathbf{p}_k),$$3

with complexity $$\mathbf{y}(\mathbf{p}_0) = \sum_{k=1}^{K} \mathbf{w}_k \cdot \mathbf{x}(\mathbf{p}_0 + \mathbf{p}_k),$$4 for $$\mathbf{y}(\mathbf{p}_0) = \sum_{k=1}^{K} \mathbf{w}_k \cdot \mathbf{x}(\mathbf{p}_0 + \mathbf{p}_k),$$5 spatial positions. RAD-Conv preserves convolutional structure and weight sharing while allowing each integration region to expand adaptively up to image dimensions. The manuscript argues that the effective window size of deformable convolution is analytically bounded by $$\mathbf{y}(\mathbf{p}_0) = \sum_{k=1}^{K} \mathbf{w}_k \cdot \mathbf{x}(\mathbf{p}_0 + \mathbf{p}_k),$$6, since each of the $$\mathbf{y}(\mathbf{p}_0) = \sum_{k=1}^{K} \mathbf{w}_k \cdot \mathbf{x}(\mathbf{p}_0 + \mathbf{p}_k),$$7 deformed points interpolates from four pixels, whereas RAD-Conv’s effective window can be adaptively bounded up to the whole feature map. For a layer with $$\mathbf{y}(\mathbf{p}_0) = \sum_{k=1}^{K} \mathbf{w}_k \cdot \mathbf{x}(\mathbf{p}_0 + \mathbf{p}_k),$$8 kernel elements, $$\mathbf{y}(\mathbf{p}_0) = \sum_{k=1}^{K} \mathbf{w}_k \cdot \mathbf{x}(\mathbf{p}_0 + \mathbf{p}_k),$$9 output positions, and region side length $\mathcal R=\{\mathbf p_k\}_{k=1}^K$0, RAD-Conv complexity is stated as

$\mathcal R=\{\mathbf p_k\}_{k=1}^K$1

which is more favorable than global attention when $\mathcal R=\{\mathbf p_k\}_{k=1}^K$2.

4. Prediction branches, exact integration, and differentiability

The implementation described in the paper mirrors deformable convolution in using companion prediction branches. A boundary-offset branch produces $\mathcal R=\{\mathbf p_k\}_{k=1}^K$3, or $\mathcal R=\{\mathbf p_k\}_{k=1}^K$4 channels in the grouped setting, and a modulation branch produces $\mathcal R=\{\mathbf p_k\}_{k=1}^K$5, or $\mathcal R=\{\mathbf p_k\}_{k=1}^K$6 channels for grouped operation. For a $\mathcal R=\{\mathbf p_k\}_{k=1}^K$7 kernel, $\mathcal R=\{\mathbf p_k\}_{k=1}^K$8, so each output location predicts 36 boundary values and 9 modulation values in the simplest configuration (Maleki et al., 18 Sep 2025).

Positivity of the offsets is enforced, following the FCOS analogy, via an exponential-style transformation. The paper does not provide a complete standalone layer diagram with all nonlinearities and kernel sizes of the predictor, but it states that the intended design is a companion convolutional layer that predicts offsets and modulation parameters from the current feature map, analogous in spirit to deformable-convolution offset heads. Grouped computation is explicitly discussed in the style of DCNv3.

A notable design choice is that rectangular aggregation is not implemented as integer-bin pooling. The manuscript instead specifies analytic computation following Precise RoI Pooling, so the operator integrates exactly over the bilinearly interpolated continuous feature surface. This has a direct optimization consequence: the region average depends continuously on the four boundaries, and gradients can therefore be propagated end-to-end through the region geometry. The paper gives the derivative with respect to the left boundary and interprets it as the sum of a normalization-area term and a boundary-crossing feature term; analogous expressions hold for the other three boundaries. The intended benefit is smooth, geometrically meaningful gradients rather than gradient behavior tied only to interpolation around a single point sample.

Architecturally, RAD-Conv is presented as a drop-in convolutional operator compatible with standard CNN pipelines. It can replace standard convolutions and deformable convolutions, and it can be used with small kernels such as $\mathcal R=\{\mathbf p_k\}_{k=1}^K$9 while still enabling large adaptive receptive fields. At the same time, the manuscript does not specify exact insertion points in named backbones, detailed optimizer settings, initialization schedules beyond positive offset constraint, or optimized kernel-level engineering for GPUs.

5. Relation to adjacent methods and terminological disambiguation

Several nearby methods use similar vocabulary but implement different adaptive mechanisms. The distinctions are substantive rather than nominal (Shi et al., 2021, Chen et al., 2020, Xiong et al., 2019, Gao et al., 2019).

Method	Adaptive primitive	Relation to RAD-Conv
RAD-Conv	Rectangular region integration with four boundary offsets per kernel element	Main subject
Reference-oriented deformable convolution in RADN	Offsets predicted from the reference feature map and applied to both branches in FR-IQA	Not named RAD-Conv
DRConv	Hard region routing and dynamic filter assignment	Changes which filter is used, not where or over what region sampling occurs
Deformable Filter Convolution	Continuous interpolation of a 3D filter at observed point offsets	Geometry-adaptive point-cloud operator, not rectangular region integration
Deformable Kernels	Kernel-space resampling to adapt the effective receptive field	Changes kernel sampling rather than image-space region aggregation

The 2021 paper “Region-Adaptive Deformable Network for Image Quality Assessment” does not define a module called Region-Aware Deformable Convolution. Its closest component is “reference-oriented deformable convolution,” a paired-reference IQA module in which offsets are predicted from the reference feature map and then applied to both the reference and distorted branches. The “region-adaptive” character of that work arises from the combination of patch-based processing, patch-level attention, and weighted patch-score aggregation rather than from a module formally named RAD-Conv (Shi et al., 2021).

Dynamic Region-Aware Convolution (DRConv) is also distinct. Its mechanism is region-wise filter routing: a learnable guided mask assigns each spatial position to one of $1\times1$0 learned regions, and the layer dynamically selects a region-specific filter for that position. The spatial sampling grid remains that of standard convolution. In consequence, DRConv adapts which filter is applied at each location, whereas RAD-Conv adapts the spatial extent and aspect ratio of the contributing region itself (Chen et al., 2020).

Other related operators illuminate further contrasts. Deformable Filter Convolution for point clouds continuously evaluates a 3D filter at exact point offsets using tri-linear interpolation, so its deformation is geometry-driven rather than boundary-regressed or region-conditioned (Xiong et al., 2019). Deformable Kernels alter kernel-space sampling to reshape the effective receptive field while leaving image-space support unchanged (Gao et al., 2019). Hardware work on deformable convolution accelerators addresses irregular memory access and bilinear interpolation for generic DCN-style sampling, but does not discuss explicit region-aware formulations such as RAD-Conv (Xu et al., 2021).

6. Conceptual contributions, current status, and limitations

The primary contribution of RAD-Conv is conceptual and mathematical rather than empirical. The manuscript proposes a new deformable-convolution primitive in which each kernel element aggregates over a learned rectangular region instead of a learned offset point. It therefore generalizes standard convolution and deformable convolution in a specific sense: standard convolution is recovered when each region collapses to fixed kernel-point locations with minimal extent and fixed modulation; deformable convolution can be interpreted as a degenerate sparse continuous-sampling case in which each kernel element samples a single shifted point reconstructed from four neighbors; and RAD-Conv extends both by permitting aggregation over a continuous, input-adaptive rectangle (Maleki et al., 18 Sep 2025).

The formulation suggests several prospective advantages. It provides a richer geometric primitive than point-based deformable sampling, decouples receptive-field geometry from kernel size, allows independent control of width and height, supports smooth differentiation through exact continuous integration, and preserves convolutional structure and weight sharing rather than replacing them with full attention. A plausible implication is that a single RAD-Conv layer can mix local and nonlocal behavior by using small rectangles where detail matters and large rectangles where long-range context is informative.

The current status of the method is explicitly pre-empirical. The paper does not contain an experimental results section with datasets, benchmarks, quantitative tables, visualizations, or ablations. No results are reported on ImageNet, COCO, ADE20K, or any other task. There are no measured comparisons against deformable convolution, attention modules, or large-kernel CNNs, and no ablations on kernel size, region parameterization, runtime, FLOPs, or memory beyond conceptual complexity discussion. The available figures are conceptual rather than empirical: they illustrate that each kernel element can define its own rectangular integration region and that these regions can vary in extent and aspect ratio across positions. The conclusion states that future work will empirically validate these capabilities across vision tasks.

The limitations are correspondingly clear from the formulation. The regions are restricted to axis-aligned rectangles, which is more flexible than fixed kernels or quadrilateral interpolation stencils but less expressive than arbitrary shapes or fully adaptive attention weights over all positions. The operator is also computationally heavier than sampling a few points, and the manuscript itself notes that RAD-Conv increases per-layer computation relative to deformable convolution. The conjecture that one RAD-Conv layer may reduce the network depth needed for large receptive fields remains unverified in the absence of experiments. Because the method averages over rectangular regions, it may smooth fine spatial structure inside a region unless modulation and learned kernel composition compensate for that effect. Practical engineering questions, including optimized kernels, memory bandwidth, and exact GPU implementation cost, are left unresolved.