Paper · 2022-02-15 · cs.CV, cs.LG
DualConv: Dual Convolutional Kernels for Lightweight Deep Neural Networks
DualConv replaces a standard 3x3 convolution with a 3x3 group convolution plus a full 1x1 convolution, summed. The group count G sets the trade-off: FLOPs fall towards 1/G + 1/9 of a standard layer. Author-reported: VGG-16 on CIFAR-10 with G=4 reaches 94.14% vs 93.95% baseline with over 60% fewer FLOPs.
Grouped convolutionDepthwise separable convolutionEfficient vision backbonesConvolutional neural networkconvolutionlightweight-cnnimage-classificationobject-detection
AI-compiled by claude-haiku-5-5 on 2026-11-13, not yet reviewed by a human. Every claim cites its source; check it when a detail matters.
In plain words
Image networks spend most of their time in convolution layers, where each output value is a weighted sum over a small patch of the input. Making those layers cheaper usually means pruning the trained network or redesigning the whole architecture. DualConv instead changes the building block itself, so it can replace the standard convolution in an existing model.
Each DualConv layer runs two filters side by side on the same input. One is a 3×3 filter that reads only a slice of the input channels, which is cheap. The other is a 1×1 filter that reads every input channel, which keeps the full information. The two outputs are added. A number G sets how the filters are split into groups: a larger G means less computation and fewer parameters, and at large G the layer behaves more like a plain 1×1 layer.
The authors test it on classic image classifiers, on the lightweight MobileNet models and on an object detector, aiming at phones and small embedded devices. All reported gains are the authors' own, so check the setup before relying on them.
TL;DR
- DualConv combines a 3×3 group convolution with a 1×1 convolution over all input channels, and sums the two outputs §III.A.
- The number of groups G controls the trade-off between cost and accuracy; FLOPs of a layer approach 1/G + 1/9 of a standard 3×3 layer §III.A eq. 5.
- Reported on CIFAR-10/100, ImageNet and PASCAL VOC with VGG-16, ResNet-50, MobileNetV1/V2 and YOLO-V3 §IV.
- Headline: MobileNetV1 on CIFAR-100 gains 4.11 points at G=4; YOLO-V3 on PASCAL VOC gains 4.4 mAP points at G=4 with about half the FLOPs tab. IV tab. VII.
- No official code was found. Results are author-reported, and several claims in the text disagree with the authors' own tables (see Pitfalls).
Problem
Deployment on embedded and mobile hardware is limited by memory, power and compute §I. The authors list three families of remedies: connection pruning, filter pruning and quantization. They argue that compressing a trained model is costly and that designing a new efficient architecture is slow and does not reuse existing backbones §I. Their goal is an efficient filter that can replace the standard 3×3 convolution in any existing CNN §I.
Two earlier filter designs are named as the starting points. Group convolution cuts cost, but its groups do not exchange information; ShuffleNet's channel shuffle is the usual fix §II.B. Heterogeneous convolution mixes 3×3 and 1×1 kernels in a shifted arrangement, which the authors argue breaks the continuity of cross-channel integration §II.C §III.A.
Method
Overview. A DualConv layer keeps the standard layer's output shape of N channels and its input of M channels, with 3×3 kernels §III.A. The N filters are split into G groups §III.A:
- Each group holds N/G filters. Each filter has a 3×3 kernel over M/G input channels, and the 3×3 slice moves along the channel dimension from group to group, without the shifted layout of heterogeneous convolution §III.A fig. 2.
- Each filter also has a 1×1 kernel over all M input channels §III.A fig. 1(e).
- The output of the 3×3 branch and the output of the 1×1 branch are added §III.A.
- No channel shuffle is used §III.A.
The 1×1 branch carries the full input, which the authors use to preserve the original information for deeper layers §III.A. The parameter count is not written as a formula in the paper. It follows from the description above: about N·M·(K²/G + 1) weights per layer, derived here and not stated by the authors. The authors give only the FLOP equations below.
Cost equations (K = kernel size, D_o = output width and height):
FL_{SC} = D_o^2 \times K^2 \times M \times N eq. 1
FL_{DC} = D_o^2 \times K^2 \times M \times N / G + D_o^2 \times M \times N eq. 4, summed from the 3×3-plus-1×1 share eq. 2 and the 1×1-only share eq. 3
R_{DC/SC} = \frac{FL_{DC}}{FL_{SC}} = \frac{1}{G} + \frac{1}{K^2} eq. 5
For comparison, the authors give these FLOP ratios against the standard layer: depthwise separable convolution 1/N + 1/K² eq. 7; group convolution 1/G eq. 9; heterogeneous convolution (K² + P − 1)/(P·K²) with P the 3×3 ratio parameter eq. 10–eq. 11.
Algorithm (the paper gives no algorithm box; this is the forward pass as the text describes it):
# DualConv layer, stride and padding not stated [§III.A] [fig. 1(e)]
# Input: x (B, M, H, W); N filters split into G groups
# Output: y (B, N, H, W)
def dual_conv(x, weights_3x3, weights_1x1, groups):
# Grouped 3x3 branch: each group of N/G filters reads M/G input
# channels, and the channel slice moves from group to group [§III.A]
y_3x3 = grouped_conv(x, weights_3x3, groups) # (B, N, H, W)
# Full 1x1 branch: every filter reads all M input channels [§III.A]
y_1x1 = conv1x1(x, weights_1x1) # (B, N, H, W)
# Sum the two branches; no channel shuffle follows [§III.A]
return y_3x3 + y_1x1Block variants. In MobileNetV1 the authors replace every depthwise separable convolution with a DualConv §IV.B. In MobileNetV2 they replace the whole inverted residual block with one DualConv and add batch normalization and ReLU6 after it §IV.B.
Hyperparameters
Values are as written in the paper. Anything not listed is "not stated" in the paper. The group values tested are G = 2, 4, 8, 16 and 32 tab. I–tab. VII.
| Name | Value | Where | Notes |
|---|---|---|---|
| Optimizer | SGD with step learning-rate decay | §IV.A §IV.C §IV.E §IV.F | Momentum not stated |
| CIFAR-10: weight decay | 5e-4 | §IV.A | Also for MobileNet on CIFAR-10 §IV.B |
| CIFAR-10: initial learning rate | 0.1, ×0.1 every 50 epochs | §IV.A | Total epochs not stated |
| CIFAR-10: VGG-16 modification | last 12 3×3 layers replaced; kernel counts unchanged | §IV.A | One fully connected layer on CIFAR-10 §IV.C |
| CIFAR-10: ResNet-50 modification | all 3×3 stride-1 convs replaced, except the first layer | §IV.A | Stride-2 convs kept on CIFAR-10 |
| CIFAR-10: MobileNetV1 modification | all depthwise separable convs replaced; first depthwise stride set to 1 | §IV.B | Stride 2 restored for ImageNet §IV.F |
| CIFAR-10: MobileNetV2 modification | inverted residual block replaced; stride 1; BN and ReLU6 added | §IV.B | |
| CIFAR-100: weight decay | 5e-4 | §IV.C | |
| CIFAR-100: initial learning rate | 0.1, ×0.2 every 60 epochs | §IV.C | Differs from CIFAR-10 schedule |
| CIFAR-100: VGG-16 head | three fully connected layers | §IV.C | |
| CIFAR-100: MobileNet settings | same as CIFAR-100 VGG/ResNet settings | §IV.D | |
| ImageNet: weight decay | 1e-4 | §IV.E | |
| ImageNet: batch size | 128 (VGG-16 and ResNet-50) | §IV.E | Not stated for MobileNets |
| ImageNet: initial learning rate | 0.01 (VGG-16), 0.1 (ResNet-50), ×0.1 every 30 epochs | §IV.E | |
| ImageNet: epochs | 90 (VGG-16 and ResNet-50) | §IV.E | |
| ImageNet: ResNet-50 modification | 3×3 convs with stride 1 and 2 replaced, except the first layer | §IV.E | Stride-2 replacement differs from YOLO-V3 |
| ImageNet: MobileNetV2 | weight decay 4e-5; initial LR 0.05; cosine decay; 150 epochs | §IV.F | |
| ImageNet: MobileNetV1 | ResNet-50 ImageNet settings | §IV.F | |
| YOLO-V3: first conv layer | not modified | §IV.I | |
| YOLO-V3: replaced convs | 3×3 stride-1 convs; stride-2 convs kept | §IV.I | |
| YOLO-V3: input size | 416×416 | §IV.I | |
| YOLO-V3: training | from scratch, 100 epochs, VOC 2007+2012 trainval (16551 images) | §IV.I | Optimizer, LR, batch size not stated |
| YOLO-V3: evaluation | VOC 2007 test, mAP as mean of per-class AP (area under PR curve) | §IV.I | |
| Framework (CIFAR-10 runs) | PyTorch; convolutions via im2col | §IV.A | |
| Timing hardware | GPU: one NVIDIA V100; CPU: one thread, Intel Core i7-8700 | §IV.G | |
| Padding, bias, BN, activation (VGG/ResNet/YOLO) | not stated | — | |
| Data augmentation | not stated | — | |
| Number of runs / seeds | not stated | — | See Pitfalls |
Results
Author-reported. Setup: CIFAR-10 and CIFAR-100 with VGG-16, ResNet-50, MobileNetV1/V2; ImageNet with VGG-16, ResNet-50, MobileNetV1/V2; PASCAL VOC 2007 test for YOLO-V3 trained from scratch. Baselines are the standard model of each architecture tab. I–tab. VII. Format: accuracy, FLOPs, parameters.
| Setting | Metric | Baseline | This paper (G) | Source |
|---|---|---|---|---|
| VGG-16, CIFAR-10 | Acc | 93.95%; 313.21M; 14.73M | G=4: 94.14%; 114.24M; 5.33M | tab. I |
| ResNet-50, CIFAR-10 | Acc | 94.08%; 1.30G; 23.52M | G=8: 94.33%; 922.99M; 17.24M | tab. I |
| MobileNetV1, CIFAR-10 | Acc | 91.91%; 46.37M; 3.22M | G=4: 93.14%; 144.03M; 10.23M | tab. II |
| MobileNetV2, CIFAR-10 | Acc | 91.99%; 64.96M; 2.37M | G=32: 90.83%; 22.42M; 786.45K | tab. II |
| VGG-16, CIFAR-100 | Acc | 72.41%; 332.48M; 34.01M | G=2: 73.04%; 211.37M; 28.29M | tab. III |
| ResNet-50, CIFAR-100 | Acc | 78.55%; 1.30G; 23.70M | G=4: 78.53%; 984.51M; 18.45M | tab. III |
| MobileNetV1, CIFAR-100 | Acc | 68.13%; 46.46M; 3.32M | G=4: 72.24%; 144.12M; 10.32M | tab. IV |
| MobileNetV2, CIFAR-100 | Acc | 68.54%; 64.96M; 2.37M | G=4: 67.86%; 31.07M; 1.09M | tab. IV |
| VGG-16, ImageNet | Top-1 / Top-5 | 72.69 / 90.91; 15.47G; 138.36M | G=2: 72.21 / 90.74; 9.54G; 132.64M | tab. V |
| ResNet-50, ImageNet | Top-1 / Top-5 | 74.27 / 92.09; 4.09G; 25.56M | G=2: 74.09 / 91.80; 3.37G; 21.16M | tab. V |
| MobileNetV1, ImageNet | Top-1 / Top-5 | 70.65 / 89.69; 568M; 4.23M | G=2: 72.29 / 90.80; 2.98G; 18.31M | tab. VI |
| MobileNetV2, ImageNet | Top-1 / Top-5 | 69.22 / 88.93; 300.79M; 3.50M | G=2: 65.45 / 84.40; 221.46M; 2.67M | tab. VI |
| YOLO-V3, PASCAL VOC 2007 | mAP | 41.24%; 32.71G; 61.63M; 26.66 ms | G=4: 45.64%; 16.41G; 30.05M; 19.79 ms | tab. VII |
Timing (ImageNet, CPU, one thread): VGG-16 falls from 307.96 ms to 148.61 ms at G=16; ResNet-50 falls from 99.21 ms to 75.19 ms at G=8 tab. V §IV.G. The authors state that GPU timing did not separate the models §IV.G.
Paper vs code
No official code.
Search, in the order the skill requires: the paper and its arXiv abstract page list no repository or project page §I; a GitHub repository search for "DualConv" returned one repository whose README describes itself as the official implementation and cites the three authors. That repository is ChipsGuardian/DualConv (AGPL-3.0). Its owner is a personal account with no name, it is not linked from the paper or arXiv, and it therefore cannot be verified as the authors'. It was not read, cited or used here. The remaining search hits were unrelated or third-party reimplementations.
Pitfalls
- Text and table disagree on VGG-16 with CIFAR-100. The text says accuracy improves as G grows to 8 §IV.C, but the table shows the best value at G=2 and a steady decline after it tab. III. Trust the table.
- Stride handling differs between experiments. On CIFAR-10 ResNet-50, only stride-1 convs are replaced §IV.A. On ImageNet ResNet-50, stride-1 and stride-2 convs are replaced §IV.E. On YOLO-V3, stride-2 convs are kept because the authors argue a stride-2 1×1 would skip input values §IV.I. The ImageNet ResNet choice is not justified in the text. "First layer" is not defined for ImageNet.
- Unstated implementation details. Padding, bias, batch-norm and activation placement for VGG/ResNet/YOLO are not stated. Only MobileNetV2 adds batch norm and ReLU6 §IV.B.
- Parameter savings on ImageNet VGG-16 are small. The fully connected layers hold about 102M parameters, so the 138.36M → 132.64M change is a small fraction of the model §IV.E tab. V.
- Abstract numbers use G=4 for MobileNetV2. The "54% fewer parameters" claim matches G=4 (1.09M vs 2.37M) tab. IV. The same 0.68-point accuracy drop also appears at G=2 (1.45M parameters), which is a smaller reduction.
- ImageNet MobileNetV2 accuracy falls clearly. The best DualConv row is 3.77 Top-1 points below baseline at G=2 and 13.25 points at G=32 tab. VI. The authors give a possible cause (the whole inverted residual block is replaced, so the features are activated once instead of twice) but do not test it §IV.F.
- Unverified HetConv gap. The text says the reproduced HetConv is 0.2–0.76% worse than the originally reported HetConv §IV.A. The originally reported figures are not in the published tables, so this gap cannot be checked from the paper's own tables.
- Reproduced baselines are the authors' own. GroupConv and HetConv were reimplemented in PyTorch with im2col by the authors, to make them comparable §IV.A. Their numbers are not the original papers' numbers.
- Run count and variance. The text does not state how many runs were made or what seeds were used. A sentence saying means over 10 runs was commented out in the source. Small gaps (for example 0.02 Top-1 points for ResNet-50 at G=4 on CIFAR-100 §IV.C) should not be read as reliable.
- Accuracy gains are absolute points. Gains such as 4.11% and 4.4% are differences in percentage points of accuracy or mAP, not relative improvements §IV.E §IV.I.
- FLOP speedup is a count, not a measurement. The 8–9× figure follows from the FLOP ratio as G grows large with K = 3 eq. 5. Measured CPU speed-ups are smaller and non-monotonic: the best CPU time is at G=16 for VGG-16 and G=8 for ResNet-50, not at the most accurate setting tab. V.
- Visual analysis is qualitative. The Grad-CAM comparison has no metric §IV.H fig. 3.
- Figures are described, not reproduced. Figure 1 compares five filter designs and Figure 2 shows the DualConv layout. Figure 3 is a heat-map comparison fig. 1 fig. 2 fig. 3.
Lineage
- Builds on: group convolution (first used in AlexNet; ShuffleNet's channel shuffle improves information exchange between groups) §II.B; heterogeneous convolution (HetConv, Singh et al., 2019; reference key singh2019hetconv) §II.C. DualConv keeps the grouped 3×3 and drops the channel shuffle §III.A.
- Compared against: depthwise separable convolution in MobileNetV1 (Howard et al., 2017) §II.A tab. II tab. IV tab. VI; filter and connection pruning methods on VGG-16 and ResNet-50 (Li et al., ThiNet, CP, NISP, AFP, SBP, RNP) tab. I tab. V.
- Superseded by: not stated.
Source and license
Adapted from “DualConv: Dual Convolutional Kernels for Lightweight Deep Neural Networks” by Jiachen Zhong, Junying Chen, Ajmal Mian (arXiv:2202.07481), used under CC BY 4.0. Changes: condensed and restructured for implementation, with tables, equations and algorithms taken from the paper and its code, each cited to its source. Not affiliated with or endorsed by the authors.