---
title: Modified Atrous Spatial Pyramid Pooling
url: https://www.emergentmind.com/topics/modified-atrous-spatial-pyramid-pooling-maspp
type: topic
---

# Modified Atrous Spatial Pyramid Pooling

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 [1606.00915][1706.05587].

## 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 \(r=1\) [1606.00915]. 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 \(r\) [1606.00915].

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 \(\{6,12,18,24\}\) [1606.00915]. 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 \((6,12,18)\) at \(OS=16\) and doubled rates at \(OS=8\); all branches use batch normalization, and their outputs are concatenated and projected [1706.05587]. 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 [1802.02611].

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
$$
r^* = \frac{l-\alpha}{6s}, \qquad \alpha \approx 32,
$$
to match ASPP field of view to input crop size \(l\) and output stride \(s\), then uses branch rates \(\{r^*,2r^*,3r^*\}\) instead of fixed defaults [2307.14179]. 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 [2303.00199].

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 [2501.13129]. 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 [1912.03183]. DenseDDSSPP, proposed for road extraction, replaces the conventional ASPP stage in DeepLabV3+ with a densely connected cascade of depthwise dilated separable convolutions [2410.14836].

A third axis is **adaptive inter-branch fusion**. In remote sensing, adaptive-ASPP in FFPNet uses four atrous branches with rates \(1,6,12,24\) plus image pooling, then applies cross-scale attention before concatenation [2008.08775]. SPAP, developed for GANs and CycleGANs, retains the atrous pyramid idea but replaces static concatenation-\(1\times1\) fusion with pairwise cascade attention and a residual blend with the input [1901.06322].

A fourth axis is **operator or geometry substitution**. DeepLabv3+ uses atrous separable convolution inside ASPP [1802.02611]. ASPDC replaces atrous branches with deformable-convolution branches at multiple dilation rates and fuses them by attention [2106.14336]. ASCSPP replaces square atrous branches with atrous strip convolutions for deformable linear object detection [2507.12744]. DAS-SK replaces plain atrous branches with DAS-SKConv branches and replaces global average pooling with strip pooling inside an enhanced ASPP [2602.08168].

| Modification axis | Representative module | Distinguishing change |
|---|---|---|
| Rate redesign | Resolution-aware ASPP [2307.14179] | \(r^*=(l-32)/(6s)\) with branches \(\{r,2r,3r\}\) |
| Dynamic scheduling | DMSA ASPP [2303.00199] | Epoch-dependent dilation table |
| Repetition / cascade | Repeated ASPP, WASP, DenseDDSSPP [2501.13129][1912.03183][2410.14836] | Three bottleneck ASPPs; waterfall cascade; dense cascaded separable pyramid |
| Adaptive fusion | adaptive-ASPP, SPAP [2008.08775][1901.06322] | Cross-scale attention; cascade spatial attention with residual blend |
| Operator substitution | atrous separable, deformable, strip, DAS-SKConv [1802.02611][2106.14336][2507.12744][2602.08168] | 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 [2602.08168][2507.12744].

## 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 [1706.05587][1802.02611]. 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 \(3\times3\) atrous branches with dilation and padding values \(6\), \(12\), and \(18\), followed by concatenation and a \(1\times1\) projection; there is no explicit image-level pooling branch in the depicted module [2103.06419]. 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 [2501.13129].

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
$$
\mathrm{attn}(Q,K,V)=\mathrm{softmax}\!\left(\frac{QK^T}{\sqrt{d_k}}\right)V,
$$
and the ASPP module is applied to strengthen feature extraction on the attention map rather than on a conventional CNN stage [2303.00199]. 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 \(x_5\) 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 [2008.08775]. 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 [2410.14836]. DAS-SK goes further: its enhanced ASPP takes a fused \(960\)-channel dual-backbone tensor, applies one \(1\times1\) branch, six DAS-SKConv branches with dilations \(\{4,8,12,18,22,26\}\), and one strip-pooling branch, concatenates them into \(1280\) channels, and projects back to \(256\) channels via \(1\times1\) convolution, batch normalization, ReLU, and dropout [2602.08168].

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 [1809.02791]. Electron micrograph denoising uses a modified Xception encoder, an ASPP module explicitly described as Google’s ASPP “without a pre-pooling \(1\times1\) convolution,” and a multi-stage decoder [1807.11234]. 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 [2106.14336].

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 \(\{r,2r,3r\}\), the ASPP center-to-center field of view in image coordinates is \(6rs\), and the approximate end-to-end field of view is \(6rs+\alpha\) with \(\alpha\approx32\). Setting this approximately equal to crop size \(l\) yields the practical rule
$$
6rs+\alpha \approx l, \qquad r^*=\frac{l-\alpha}{6s},
$$
which turns fixed-rate ASPP into a resolution-aware MASPP without changing branch count [2307.14179].

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:
$$
R_{\mathrm{List}}(x)=
\begin{cases}
[1,1,2,3], & x \bmod 10 \in \{1,3,5,7\},\\
[1,1,3,5], & x \bmod 10 \in \{2,4,6\},\\
[1,3,6,9], & x \bmod 10 \in \{8,9\},\\
[1,6,12,18], & x \bmod 10 = 0.
\end{cases}
$$
This is stage-dependent in training but not input-dependent, which distinguishes it from content-adaptive ASPP or dynamic-convolution variants [2303.00199].

The third idea is that branch interaction can be made explicit. In adaptive-ASPP, each scale feature \(X_i\) interacts with other scale features \(X_j\) through cross-scale attention:
$$
A_{ji}=\sigma_2\Big(\Psi^T\big(\sigma_1(W_x^T * X_i + W_g^T * X_j)\big)\Big), \qquad
\tilde{X}_i = \sum_{j\neq i} A_{ji} X_i + X_i.
$$
Branch fusion is therefore no longer passive concatenation; each scale is modulated by the others before final aggregation [2008.08775]. SPAP applies a related principle in generative models, using attentive pairwise fusion and a residual mixture
$$
y_{out}=\gamma\,o+(1-\gamma)\,x_{in},
$$
so the pyramid output is blended with the original feature map rather than replacing it outright [1901.06322].

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 \(\Delta p\) and modulations \(\Delta m\), then fuses branch outputs by spatial attention
$$
f_o= \sum_{i = 1}^{4} a_i * f_i,\qquad \sum_{i = 1}^{4} a_{ij} = 1.
$$
This converts the usual “same-rate, same-grid, different receptive field” logic into “different receptive field plus learned sampling geometry” [2106.14336]. 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 [2602.08168].

## 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 \(8\), the guideline predicts \(r^*\approx15\) for \(769\times769\) crops; mIoU improves from \(79.33\) to \(79.93\) for ResNet-50 and from \(78.88\) to \(79.90\) for ResNet-101 when moving from the default \(r=12\) to \(r=15\). On iSAID with \(896\times896\) input, \(r=18\) improves mIoU from \(66.28\) to \(67.03\) over the default \(r=12\) [2307.14179]. 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 \(90.37\%\) to \(90.91\%\) and mIoU from \(79.96\%\) to \(81.33\%\), whereas standard ASPP slightly underperforms the feature-fusion pyramid without ASPP adaptation [2008.08775]. 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 \(80.22\%\) mIoU with \(47.482\)M parameters, while DeepLab reaches \(76.35\%\) with \(59.869\)M parameters; the paper reports a \(20.69\%\) parameter reduction. On Cityscapes validation, WASPnet reaches \(74.0\%\) mIoU versus \(71.0\%\) for DeepLab [1912.03183]. 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 \(65.92\%\) to \(67.21\%\) and \(F_1\) from \(75.60\%\) to \(79.29\%\) relative to original DeepLabV3+; on DeepGlobe Roads it improves IoU from \(69.05\%\) to \(71.61\%\) and \(F_1\) from \(77.96\%\) to \(81.75\%\) [2410.14836]. ASCSPP, a lighter geometry-aware replacement, improves mIoU from \(0.4386\) to \(0.5069\) over ASPP while reducing speed only from \(223\) FPS to \(217\) FPS in the reported pyramid-module comparison [2507.12744].

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 \(93.85\pm1.04\) to \(94.20\pm0.51\) over ResU-Net on LiTS17, and the final SAR-U-Net reaches \(95.71\pm0.55\) Dice [2106.14336]. 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 [2501.13129]. DMSA reports overall gains over TransFGU on COCO 80 from \(12.69\) to \(14.69\) mIoU and from \(64.31\) to \(69.70\) accuracy, yet it does not isolate standard fixed-rate ASPP versus its epoch-cycled variant [2303.00199].

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 [2303.00199]. The repeated-ASPP Attention U-Net describes “Attention based UNet with three ASPP blocks” rather than a new operator [2501.13129]. 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 [2410.17863]. 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 [1802.02611]. Resolution-aware ASPP changes only the base rate \(r\) [2307.14179]. DMSA changes the dilation schedule across epochs and relocates ASPP to a ViT attention pathway, but does not specify a new branch layout [2303.00199]. Repeated ASPP changes deployment count and bottleneck placement, not necessarily operator internals [2501.13129].

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 [2303.00199]. The repeated-ASPP brain-tumor paper does not state branch number, exact dilation rates, or fusion operator for each ASPP block [2501.13129]. DenseDDSSPP does not provide a complete layer-by-layer specification of all dense pyramid stages [2410.14836]. ASCSPP states that varying dilation rates are used but does not list their exact numeric values in the provided text [2507.12744]. Even in stronger papers, key details such as strip-pooling kernel geometry or dropout probability may be absent [2602.08168].

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 [2303.00199]. The repeated-ASPP Attention U-Net does not isolate one ASPP block against three [2501.13129]. DenseDDSSPP is compared against complete baselines, yet the contribution of dense cascading versus SE integration or backbone choice is not fully disentangled [2410.14836]. 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.

Source: https://www.emergentmind.com/topics/modified-atrous-spatial-pyramid-pooling-maspp