Modified Atrous Spatial Pyramid Pooling

• Modified ASPP is a design pattern that reconfigures standard ASPP by altering dilation policies and branch structures to capture multi-scale contextual features.
• Research shows that adjustments like dynamic rate adaptation, cascading blocks, and adaptive fusion can boost segmentation accuracy and efficiency.
• Empirical evidence across domains demonstrates that tailored MASPP designs improve key metrics such as mIoU and Dice scores through operator substitutions and refined fusion.

Modified Atrous Spatial Pyramid Pooling, or MASPP, is best understood as an Editor’s term for the family of modules that modify the original Atrous Spatial Pyramid Pooling (ASPP) idea rather than as a single standardized operator. In the cited literature, many papers do not explicitly name their block “MASPP”; instead they retain the label ASPP or introduce alternatives such as “adaptive-ASPP,” “enhanced ASPP,” “ASCSPP,” “WASP,” or repeated ASPP. What unifies these designs is the preservation of ASPP’s central function—multi-scale context extraction through atrous sampling—while changing dilation policy, branch topology, fusion, operator choice, or insertion point within a larger network (Chen et al., 2016, Chen et al., 2017).

1. Canonical ASPP and the baseline from DeepLab

The canonical starting point is DeepLab’s use of atrous convolution to enlarge receptive field without reducing feature-map resolution. In the original formulation, atrous convolution is defined as

$y[i] = \sum_{k=1}^{K} x[i + r \cdot k]\, w[k],$

where $r$ is the atrous rate; standard convolution is recovered when $r=1$ (Chen et al., 2016). The effective kernel size is

$k_{\text{eff}} = k + (k-1)(r-1),$

so a $3\times3$ kernel can realize substantially larger fields of view by increasing $r$ (Chen et al., 2016).

DeepLab’s original ASPP probes a single feature map with parallel atrous branches at multiple rates. In the VGG-based setting, ASPP-S used four parallel branches with rates $\{2,4,8,12\}$, while ASPP-L used $\{6,12,18,24\}$ (Chen et al., 2016). DeepLabv3 then redefined the mainstream ASPP template more explicitly: one $1\times1$ convolution branch, three $3\times3$ atrous branches, and one image-level pooling branch, with rates $r$0 at $r$1 and doubled rates at $r$2; all branches use batch normalization, and their outputs are concatenated and projected (Chen et al., 2017). DeepLabv3+ preserved this encoder-side ASPP role while replacing standard convolutions inside ASPP and the decoder with atrous separable convolutions in the Xception-based variant (Chen et al., 2018).

This baseline matters because later MASPP-like proposals are usually modifications of one of these two DeepLab lineages: either the earlier “parallel multi-rate atrous branches” formulation or the later “parallel branches plus image-level pooling and projection” formulation.

2. Principal modification axes

The literature shows several recurring ways in which ASPP is modified. Some works alter only the dilation policy, some repeat or cascade ASPP blocks, and others replace the branch operator itself or make branch interaction adaptive.

A first axis is rate adaptation. One line keeps the ASPP topology fixed but changes how rates are chosen. The resolution-aware design in “Resolution-Aware Design of Atrous Rates for Semantic Segmentation Networks” derives

$r$3

to match ASPP field of view to input crop size $r$4 and output stride $r$5, then uses branch rates $r$6 instead of fixed defaults (Kim et al., 2023). DMSA follows a different strategy: it inserts ASPP into a ViT attention pathway and cycles dilation schedules across epochs rather than predicting them from image content (Yang et al., 2023).

A second axis is structural repetition or cascading. The brain-tumor segmentation model based on Attention U-Net places three ASPP blocks at the bottleneck; the modification is repetition of standard ASPP usage rather than internal redesign (Chowdhury et al., 22 Jan 2025). WASP, or Waterfall Atrous Spatial Pooling, replaces fully parallel ASPP with a progressive cascade that retains intermediate atrous outputs for final fusion, preserving multi-scale fields of view with fewer parameters (Artacho et al., 2019). DenseDDSSPP, proposed for road extraction, replaces the conventional ASPP stage in DeepLabV3+ with a densely connected cascade of depthwise dilated separable convolutions (Mahara et al., 2024).

A third axis is adaptive inter-branch fusion. In remote sensing, adaptive-ASPP in FFPNet uses four atrous branches with rates $r$7 plus image pooling, then applies cross-scale attention before concatenation (Xu et al., 2020). SPAP, developed for GANs and CycleGANs, retains the atrous pyramid idea but replaces static concatenation-$r$8 fusion with pairwise cascade attention and a residual blend with the input (Sun et al., 2019).

A fourth axis is operator or geometry substitution. DeepLabv3+ uses atrous separable convolution inside ASPP (Chen et al., 2018). ASPDC replaces atrous branches with deformable-convolution branches at multiple dilation rates and fuses them by attention (Huo et al., 2021). ASCSPP replaces square atrous branches with atrous strip convolutions for deformable linear object detection (Liu et al., 17 Jul 2025). DAS-SK replaces plain atrous branches with DAS-SKConv branches and replaces global average pooling with strip pooling inside an enhanced ASPP (Chee et al., 9 Feb 2026).

Modification axis	Representative module	Distinguishing change
Rate redesign	Resolution-aware ASPP (Kim et al., 2023)	$r$9 with branches $r=1$0
Dynamic scheduling	DMSA ASPP (Yang et al., 2023)	Epoch-dependent dilation table
Repetition / cascade	Repeated ASPP, WASP, DenseDDSSPP (Chowdhury et al., 22 Jan 2025, Artacho et al., 2019, Mahara et al., 2024)	Three bottleneck ASPPs; waterfall cascade; dense cascaded separable pyramid
Adaptive fusion	adaptive-ASPP, SPAP (Xu et al., 2020, Sun et al., 2019)	Cross-scale attention; cascade spatial attention with residual blend
Operator substitution	atrous separable, deformable, strip, DAS-SKConv (Chen et al., 2018, Huo et al., 2021, Liu et al., 17 Jul 2025, Chee et al., 9 Feb 2026)	New branch operators and pooling geometry

These axes are not mutually exclusive. Several later designs combine more than one: DAS-SK changes both branch operators and pooling geometry, while ASCSPP changes kernel geometry and also modifies how original features are reintroduced (Chee et al., 9 Feb 2026, Liu et al., 17 Jul 2025).

3. Representative architectural realizations

The architectural placement of MASPP-like modules varies considerably. In DeepLab-style segmentation, the ASPP family remains an encoder-side context block placed after deep backbone features and before decoder refinement (Chen et al., 2017, Chen et al., 2018). In U-Net derivatives, however, the same idea is often transplanted to the bottleneck or even to the output side.

SAR-U-Net is a clear example of simplified ASPP integration in a medical encoder-decoder. It places one ASPP module at the bottleneck transition layer and another at the decoder output side. Each ASPP consists of three parallel $r=1$1 atrous branches with dilation and padding values $r=1$2, $r=1$3, and $r=1$4, followed by concatenation and a $r=1$5 projection; there is no explicit image-level pooling branch in the depicted module (Wang et al., 2021). The repeated-ASPP Attention U-Net for brain-tumor segmentation also uses the bottleneck, but with three ASPP blocks rather than one, again without introducing a new named branch topology (Chowdhury et al., 22 Jan 2025).

Transformer-based use also appears. DMSA inserts an ASPP-like block after the multi-head self-attention output of a ViT encoder, then subjects it to an epoch-cycled dilation schedule. The attention itself is written as

$r=1$6

and the ASPP module is applied to strengthen feature extraction on the attention map rather than on a conventional CNN stage (Yang et al., 2023). This is a narrow but distinctive form of MASPP: the branch layout is not re-specified, but the insertion point and dilation policy are both modified.

Remote sensing provides several more elaborate realizations. FFPNet’s adaptive-ASPP sits at the top of a ResNet-101 encoder, receives the backbone output, and produces $r=1$7 for subsequent pyramid fusion. It contains four atrous branches and an image-pooling branch, but inserts cross-scale attention among the atrous branches before final concatenation (Xu et al., 2020). DenseDDSSPP, integrated into DeepLabV3+ for road extraction, is not parallel ASPP in the standard sense; instead it forms a dense cascade of depthwise dilated separable convolutions whose outputs are concatenated stage by stage (Mahara et al., 2024). DAS-SK goes further: its enhanced ASPP takes a fused $r=1$8-channel dual-backbone tensor, applies one $r=1$9 branch, six DAS-SKConv branches with dilations $k_{\text{eff}} = k + (k-1)(r-1),$0, and one strip-pooling branch, concatenates them into $k_{\text{eff}} = k + (k-1)(r-1),$1 channels, and projects back to $k_{\text{eff}} = k + (k-1)(r-1),$2 channels via $k_{\text{eff}} = k + (k-1)(r-1),$3 convolution, batch normalization, ReLU, and dropout (Chee et al., 9 Feb 2026).

Outside semantic segmentation proper, MASPP-like behavior appears in task-specific heads. DMAC for constrained image splicing detection feeds multi-level correlation maps into an ASPP-based mask generation module; here ASPP is repurposed as a multi-scale decoder over correlation features rather than standard image features (Liu et al., 2018). Electron micrograph denoising uses a modified Xception encoder, an ASPP module explicitly described as Google’s ASPP “without a pre-pooling $k_{\text{eff}} = k + (k-1)(r-1),$4 convolution,” and a multi-stage decoder (Ede, 2018). Non-uniform motion deblurring uses ASPDC, which is more accurately a deformable atrous pyramid than a literal pooling block, but it preserves the multi-rate atrous-branch logic (Huo et al., 2021).

This breadth suggests that MASPP is less a single module than a reusable multiscale design pattern whose exact semantics depend on surrounding architecture and target signal geometry.

4. Mathematical design principles

Despite their diversity, MASPP-family modules are organized around a small set of mathematical ideas. The first is that dilation rate controls image-scale field of view. The resolution-aware ASPP analysis makes this explicit: for the standard three-branch rate pattern $k_{\text{eff}} = k + (k-1)(r-1),$5, the ASPP center-to-center field of view in image coordinates is $k_{\text{eff}} = k + (k-1)(r-1),$6, and the approximate end-to-end field of view is $k_{\text{eff}} = k + (k-1)(r-1),$7 with $k_{\text{eff}} = k + (k-1)(r-1),$8. Setting this approximately equal to crop size $k_{\text{eff}} = k + (k-1)(r-1),$9 yields the practical rule

$3\times3$0

which turns fixed-rate ASPP into a resolution-aware MASPP without changing branch count (Kim et al., 2023).

The second idea is that dilation can be made dynamic over training rather than fixed over architecture definition. DMSA defines a dilation-rate return function through an epoch-modulo-10 table:

$3\times3$1

This is stage-dependent in training but not input-dependent, which distinguishes it from content-adaptive ASPP or dynamic-convolution variants (Yang et al., 2023).

The third idea is that branch interaction can be made explicit. In adaptive-ASPP, each scale feature $3\times3$2 interacts with other scale features $3\times3$3 through cross-scale attention:

$3\times3$4

Branch fusion is therefore no longer passive concatenation; each scale is modulated by the others before final aggregation (Xu et al., 2020). SPAP applies a related principle in generative models, using attentive pairwise fusion and a residual mixture

$3\times3$5

so the pyramid output is blended with the original feature map rather than replacing it outright (Sun et al., 2019).

The fourth idea is that the atrous branch operator itself can be changed. ASPDC conceptually replaces each fixed-grid atrous branch with a deformable branch that learns offsets $3\times3$6 and modulations $3\times3$7, then fuses branch outputs by spatial attention

$3\times3$8

This converts the usual “same-rate, same-grid, different receptive field” logic into “different receptive field plus learned sampling geometry” (Huo et al., 2021). DAS-SK changes the branch operator in another way: each ASPP branch is a DAS-SKConv that concatenates an atrous separable path and a standard atrous path, then applies selective-kernel attention before ASPP-level concatenation (Chee et al., 9 Feb 2026).

5. Empirical evidence across domains

Empirical support for MASPP-like modifications is substantial, but the strength of evidence varies by paper. The cleanest results come from controlled replacements of standard ASPP or fixed-rate defaults.

The resolution-aware rate design shows consistent gains over default DeepLab settings across both small-crop and large-crop regimes. On Cityscapes with DeepLabV3 and output stride $3\times3$9, the guideline predicts $r$0 for $r$1 crops; mIoU improves from $r$2 to $r$3 for ResNet-50 and from $r$4 to $r$5 for ResNet-101 when moving from the default $r$6 to $r$7. On iSAID with $r$8 input, $r$9 improves mIoU from $\{2,4,8,12\}$0 to $\{2,4,8,12\}$1 over the default $\{2,4,8,12\}$2 (Kim et al., 2023). This is one of the clearest demonstrations that a modified ASPP need not change topology at all; rate policy alone can matter.

Adaptive-ASPP in FFPNet also has a direct within-framework comparison. On the Vaihingen dataset, replacing standard ASPP with adaptive-ASPP changes overall accuracy from $\{2,4,8,12\}$3 to $\{2,4,8,12\}$4 and mIoU from $\{2,4,8,12\}$5 to $\{2,4,8,12\}$6, whereas standard ASPP slightly underperforms the feature-fusion pyramid without ASPP adaptation (Xu et al., 2020). The authors explicitly interpret this as evidence that fixed ASPP does not fuse context well enough for geometric variations in remote sensing scenes.

WASP provides one of the clearest efficiency arguments for a topology-modified ASPP. On Pascal VOC validation, WASPnet reaches $\{2,4,8,12\}$7 mIoU with $\{2,4,8,12\}$8M parameters, while DeepLab reaches $\{2,4,8,12\}$9 with $\{6,12,18,24\}$0M parameters; the paper reports a $\{6,12,18,24\}$1 parameter reduction. On Cityscapes validation, WASPnet reaches $\{6,12,18,24\}$2 mIoU versus $\{6,12,18,24\}$3 for DeepLab (Artacho et al., 2019). Here the modification is specifically the waterfall cascade that reuses intermediate features rather than recomputing full parallel branches.

Road extraction furnishes two additional examples. DenseDDSSPP, integrated into DeepLabV3+, improves the Massachusetts Roads IoU from $\{6,12,18,24\}$4 to $\{6,12,18,24\}$5 and $\{6,12,18,24\}$6 from $\{6,12,18,24\}$7 to $\{6,12,18,24\}$8 relative to original DeepLabV3+; on DeepGlobe Roads it improves IoU from $\{6,12,18,24\}$9 to $1\times1$0 and $1\times1$1 from $1\times1$2 to $1\times1$3 (Mahara et al., 2024). ASCSPP, a lighter geometry-aware replacement, improves mIoU from $1\times1$4 to $1\times1$5 over ASPP while reducing speed only from $1\times1$6 FPS to $1\times1$7 FPS in the reported pyramid-module comparison (Liu et al., 17 Jul 2025).

Medical and hybrid backbones show similar patterns, but usually at system level rather than clean module-only isolation. In SAR-U-Net, ASPP-ResU-Net improves Dice from $1\times1$8 to $1\times1$9 over ResU-Net on LiTS17, and the final SAR-U-Net reaches $3\times3$0 Dice (Huo et al., 2021). The repeated-ASPP Attention U-Net achieves the best reported scores among UNet, Attention U-Net, and Attention U-Net with SPP across T1C, T2 FLAIR, and T2W, but the paper does not isolate the gain from using three ASPP blocks instead of one (Chowdhury et al., 22 Jan 2025). DMSA reports overall gains over TransFGU on COCO 80 from $3\times3$1 to $3\times3$2 mIoU and from $3\times3$3 to $3\times3$4 accuracy, yet it does not isolate standard fixed-rate ASPP versus its epoch-cycled variant (Yang et al., 2023).

Taken together, these studies show that MASPP-like changes can improve segmentation quality, robustness to scale variation, or efficiency, but they do so through heterogeneous mechanisms: rate selection, attention, dense cascading, operator substitution, or deployment strategy.

6. Nomenclature, limitations, and reproducibility

A persistent misconception is that MASPP refers to a single well-defined module. The literature does not support that interpretation. Several papers explicitly do not introduce a separately branded “MASPP” block even when they modify ASPP in practice. DMSA describes “ASPP with dynamic dilated rate” rather than MASPP (Yang et al., 2023). The repeated-ASPP Attention U-Net describes “Attention based UNet with three ASPP blocks” rather than a new operator (Chowdhury et al., 22 Jan 2025). CASCRNet includes an ASPP block, but the available description does not show a distinct modified ASPP micro-architecture, only ASPP embedded in a custom SCR-based classifier (Srinanda et al., 2024). This suggests that MASPP is most accurately used as a comparative category rather than as a canonical name.

Another misconception is that every MASPP must redesign branch internals. Some modifications are narrow. DeepLabv3+ changes the convolutional operator to atrous separable convolution while keeping the general ASPP template (Chen et al., 2018). Resolution-aware ASPP changes only the base rate $3\times3$5 (Kim et al., 2023). DMSA changes the dilation schedule across epochs and relocates ASPP to a ViT attention pathway, but does not specify a new branch layout (Yang et al., 2023). Repeated ASPP changes deployment count and bottleneck placement, not necessarily operator internals (Chowdhury et al., 22 Jan 2025).

Reproducibility is also uneven. Several papers omit branch counts, exact dilation values, channel dimensions, or fusion operators. DMSA does not enumerate ASPP branch layout, channel widths, normalization inside ASPP, or projection details (Yang et al., 2023). The repeated-ASPP brain-tumor paper does not state branch number, exact dilation rates, or fusion operator for each ASPP block (Chowdhury et al., 22 Jan 2025). DenseDDSSPP does not provide a complete layer-by-layer specification of all dense pyramid stages (Mahara et al., 2024). ASCSPP states that varying dilation rates are used but does not list their exact numeric values in the provided text (Liu et al., 17 Jul 2025). Even in stronger papers, key details such as strip-pooling kernel geometry or dropout probability may be absent (Chee et al., 9 Feb 2026).

A further limitation is the frequent absence of clean ablations. DMSA reports ASPP-related gains but not a direct standard-ASPP versus dynamic-ASPP comparison (Yang et al., 2023). The repeated-ASPP Attention U-Net does not isolate one ASPP block against three (Chowdhury et al., 22 Jan 2025). DenseDDSSPP is compared against complete baselines, yet the contribution of dense cascading versus SE integration or backbone choice is not fully disentangled (Mahara et al., 2024). These omissions matter because “MASPP” often bundles several changes at once.

The most defensible generalization is therefore cautious: modified ASPP in contemporary arXiv literature is not one operator but a design space built around the original atrous pyramid. This suggests that the technical question is usually not whether a model “has MASPP,” but which ASPP dimension it modifies—rate selection, branch topology, fusion rule, branch operator, pooling geometry, or architectural placement—and whether the paper provides sufficiently controlled evidence for that specific change.