mlsanityData to 7 Oct 2026

Revisiting Dynamic Convolution via Matrix Decomposition

arXiv 2103.08756 ↗First posted 15 Mar 2021cs.CVcited by 35

Not on the map. Why?

Abstract (arXiv)

Recent research in dynamic convolution shows substantial performance boost for efficient CNNs, due to the adaptive aggregation of K static convolution kernels. It has two limitations: (a) it increases the number of convolutional weights by K-times, and (b) the joint optimization of dynamic attention and static convolution kernels is challenging. In this paper, we revisit it from a new perspective of matrix decomposition and reveal the key issue is that dynamic convolution applies dynamic attention over channel groups after projecting into a higher dimensional latent space. To address this issue, we propose dynamic channel fusion to replace dynamic attention over channel groups. Dynamic channel fusion not only enables significant dimension reduction of the latent space, but also mitigates the joint optimization difficulty. As a result, our method is easier to train and requires significantly fewer parameters without sacrificing accuracy. Source code is at https://github.com/liyunsheng13/dcd.

Key facts. Cited by 35 papers in the map (data through 7 Oct 2026). Builds on 29 of its 35 references that are in the map. First posted on arXiv 15 Mar 2021.

Builds on

29 of 35 references.

References · 29

  1. 2015Deep Residual Learning for Image Recognitioncited by 51,809
  2. 2019EfficientNet: Rethinking Model Scaling for Convolutional Neural Networkscited by 6,415
  3. 2017mixup: Beyond Empirical Risk Minimizationcited by 5,155
  4. 2018MobileNetV2: Inverted Residuals and Linear Bottleneckscited by 4,773
  5. 2017Squeeze-and-Excitation Networkscited by 4,763
  6. 2017MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applicationscited by 3,827
  7. 2019Searching for MobileNetV3cited by 1,984
  8. 2017ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devicescited by 1,194
  9. 2018ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Designcited by 1,132
  10. 2019EfficientDet: Scalable and Efficient Object Detectioncited by 1,018
  11. 2016HyperNetworkscited by 1,003
  12. 2019Once-for-All: Train One Network and Specialize it for Efficient Deploymentcited by 778
17 more
  1. 2019GhostNet: More Features from Cheap Operationscited by 410
  2. 2019Selective Kernel Networkscited by 401
  3. 2019Dynamic Convolution: Attention over Convolution Kernelscited by 355
  4. 2020Conditional Convolutions for Instance Segmentationcited by 314
  5. 2014Exploiting Linear Structure Within Convolutional Networks for Efficient Evaluationcited by 304
  6. 2018Slimmable Neural Networkscited by 301
  7. 2019CondConv: Conditionally Parameterized Convolutions for Efficient Inferencecited by 273
  8. 2014Speeding-up Convolutional Neural Networks Using Fine-tuned CP-Decompositioncited by 199
  9. 2015Compression of Deep Convolutional Neural Networks for Fast and Low Power Mobile Applicationscited by 161
  10. 2019MixConv: Mixed Depthwise Convolutional Kernelscited by 108
  11. 2019AdderNet: Do We Really Need Multiplications in Deep Learning?cited by 82
  12. 2020Stable Low-rank Tensor Decomposition for Compression of Convolutional Neural Networkcited by 74
  13. 2020Rethinking Bottleneck Structure for Efficient Mobile Network Designcited by 54
  14. 2020WeightNet: Revisiting the Design Space of Weight Networkscited by 50
  15. 2020Dynamic Region-Aware Convolutioncited by 49
  16. 2019Factorized Higher-Order CNNs with an Application to Spatio-Temporal Emotion Estimationcited by 40
  17. 2020Dynamic Group Convolution for Accelerating Convolutional Neural Networkscited by 15

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2021: 4; 2022: 11; 2023: 6; 2024: 6; 2025: 6; 2026: 2.

35 papers.

Sort:

By area:

  1. 2022Self-Sustaining Representation Expansion for Non-Exemplar Class-Incremental Learningcited by 118
  2. 2022Efficient and Degradation-Adaptive Network for Real-World Image Super-Resolutioncited by 96
  3. 2021Dynamic Slimmable Networkcited by 83
  4. 2022Omni-Dimensional Dynamic Convolutioncited by 55
  5. 2022Multi-stage image denoising with the wavelet transformcited by 37
  6. 2021TAda! Temporally-Adaptive Convolutions for Video Understandingcited by 36
  7. 2023Language Adaptive Weight Generation for Multi-task Visual Groundingcited by 35
  8. 2023GaitGCI: Generative Counterfactual Intervention for Gait Recognitioncited by 30
  9. 2021Dynamic Transfer for Multi-Source Domain Adaptationcited by 29
  10. 2024Multilinear Mixture of Experts: Scalable Expert Specialization through Factorizationcited by 28
  11. 2025Frequency Dynamic Convolution for Dense Image Predictioncited by 19
  12. 2025LSNet: See Large, Focus Smallcited by 15
23 more
  1. 2022Shadow-Aware Dynamic Convolution for Shadow Removalcited by 11
  2. 2022PAD-Net: An Efficient Framework for Dynamic Networkscited by 9
  3. 2023Temporally-Adaptive Models for Efficient Video Understandingcited by 6
  4. 2024KernelWarehouse: Rethinking the Design of Dynamic Convolutioncited by 6
  5. 2022FAMLP: A Frequency-Aware MLP-Like Architecture For Domain Generalizationcited by 4
  6. 2023KernelWarehouse: Towards Parameter-Efficient Dynamic Convolutioncited by 4
  7. 2022RepMode: Learning to Re-parameterize Diverse Experts for Subcellular Structure Predictioncited by 3
  8. 2023Boosting Convolutional Neural Networks with Middle Spectrum Grouped Convolutioncited by 3
  9. 2024Image Captioning via Dynamic Path Customizationcited by 3
  10. 2022SD-Conv: Towards the Parameter-Efficiency of Dynamic Convolutioncited by 2
  11. 2023Dynamic Mobile-Former: Strengthening Dynamic Convolution with Attention and Residual Connection in Kernel Spacecited by 2
  12. 2024Metric Convolutions: A Unifying Theory to Adaptive Image Convolutionscited by 2
  13. 2024Multi-Convformer: Extending Conformer with Multiple Convolution Kernelscited by 2
  14. 2021Dynamic Parameterized Network for CTR Predictioncited by 1
  15. 2022Should All Proposals be Treated Equally in Object Detection?cited by 1
  16. 2025FMDConv: Fast Multi-Attention Dynamic Convolution via Speed-Accuracy Trade-offcited by 1
  17. 2022Dual Complementary Dynamic Convolution for Image Recognitioncited by 0
  18. 2024SGDM: Static-Guided Dynamic Module Make Stronger Visual Modelscited by 0
  19. 2025Wavelet-Enhanced Desnowing: A Novel Single Image Restoration Approach for Traffic Surveillance under Adverse Weather Conditionscited by 0
  20. 2025KernelDNA: Dynamic Kernel Sharing via Decoupled Naive Adapterscited by 0
  21. 2025DISK: Differentiable Sparse Kernel Complex for Efficient Spatially-Variant Convolutioncited by 0
  22. 2026Domain Adaptation Without the Compute Burden for Efficient Whole Slide Image Analysiscited by 0
  23. 2026MCL: Meta Convolution Layercited by 0