Research map / Detection and segmentation / CNN features
MogaNet: Multi-order Gated Aggregation Network
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MogaNet, a pure ConvNet family using multi-order dilated depth-wise convolutions with gated aggregation and a channel aggregation block, motivated by multi-order game-theoretic interaction analysis, to reach strong accuracy-efficiency trade-offs on ImageNet and dense/pose/video tasks.
Problem
Modern ConvNets, despite large kernels, encode extreme low- or high-order interactions and under-use expressive middle-order interactions (a representation bottleneck), and channel MLPs are redundant and parameter-hungry.
Approach
Combine parallel dilated depth-wise convolutions that capture low, middle and high-order interactions with a SiLU gated aggregation (spatial block), plus a lightweight channel aggregation block that reallocates channel features with a channel-reducing projection, guided by multi-order interaction analysis.
Key results
- MogaNet-T reaches 80.0% top-1 on ImageNet-1K at 256^2 with 5.2M params, outperforming ParC-Net-S by 1.0 with 2.04G lower FLOPs.
- MogaNet-XL reaches 87.8% top-1 on ImageNet-1K after IN-21K pre-training with 181M params, saving 169M params versus ConvNeXt-XL.
- Cascade Mask R-CNN with MogaNet-XL achieves 56.2 AP^b on COCO, +1.4 over ConvNeXt-L and +2.3 over RepLKNet-31L.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MogaNet-S instead of Swin-T or ConvNeXt-T | ImageNet-1K classification · ImageNet-1K, 224² | top-1 accuracy: 2.1 and 1.2 | Helped |
| Use MogaNet-T instead of ParC-Net-S | ImageNet-1K classification · ImageNet-1K, 256² | top-1 accuracy: 1.0 | Helped |
| Pre-train MogaNet-XL on ImageNet-21K | MogaNet-XL · ImageNet-1K classification | top-1 accuracy | Helped |
| Use MogaNet-T instead of Swin-T | Mask R-CNN · COCO val2017 | APᵇ: 0.4 | Helped |
| Use MogaNet-XL instead of ConvNeXt-L | Cascade Mask R-CNN · COCO val2017 | APᵇ: +1.4 | Helped |
| Use MogaNet-XL instead of ConvNeXt-L | UperNet · ADE20K | mIoU: 0.3 | Helped |
| Use MogaNet blocks instead of the SimVP baseline | SimVP · MMNIST video prediction | MSE: 6.58 MSE | Helped |
| Add the channel aggregation module CA with r = 4 (ablation) | MogaNet baseline · ImageNet-1K | performance: 0.6% | Helped |
| Use CA with r = 4 instead of the baseline MLP with r = 8 (ablation) | MogaNet channel aggregation · ImageNet-1K | performance | No clear effect |
| Add the proposed spatial aggregation modules, including FD and Moga (ablation) | MogaNet-S baseline · ImageNet-1K | performance | Helped |
| Add the proposed channel aggregation module CA (ablation) | MogaNet-S baseline · ImageNet-1K | performance | Helped |
| Use SiLU in both Moga branches (ablation) | MogaNet · ImageNet-1K | performance | Helped |
Surprising findings
- MogaNet-XL reaches 85.1% at 224^2 without pre-training, improving ConvNeXt-L by 0.8, suggesting easier convergence.
Limitations
- Gains on some downstream tasks are small relative to parameter increases in parts of the comparison
- Interaction analysis relies on masking-based approximations and is not a direct causal explanation of accuracy gains
Experimental setup
Datasets: ImageNet-1K ImageNet-21K COCO ADE20K FFHQ FreiHAND Moving MNIST Stirling/ESRC 3D
Benchmarks: ImageNet-1K COCO ADE20K COCO pose Stirling/ESRC 3D FreiHAND Moving MNIST
Metrics: top-1 accuracy AP^b AP^m mIoU AP 3DRMSE PA-MPJPE MSE SSIM FLOPs
Baselines: ConvNeXt Swin Transformer DeiT ParC-Net HorNet SLaK RepLKNet RegNetY CoAtNet Uniformer
Builds on: ConvNeXt multi-order game-theoretic interaction MetaFormer SimVP
Closest papers by meaning
- ConvNeXt V2: Co-Designing and Scaling ConvNets With Masked Autoencoders (CVPR 2023 · similarity 0.87)
- A ConvNet for the 2020s (CVPR 2022 · similarity 0.87)
- More ConvNets in the 2020s: Scaling up Kernels Beyond 51x51 using Sparsity (ICLR 2023 · similarity 0.86)
- FasterViT: Fast Vision Transformers with Hierarchical Attention (ICLR 2024 · similarity 0.85)
- DECO: Unleashing the Potential of ConvNets for Query-based Detection and Segmentation (ICLR 2025 · similarity 0.85)
- LSNet: See Large, Focus Small (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.