Research map / Detection and segmentation / CNN features
Dynamic Clustering Convolutional Neural 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 DCCNeXt, a CNN whose Dynamic Clustering Convolution groups patches into global clusters via top-K ℓ2 distance on subvectors and convolves each cluster with shared kernels, reporting ImageNet, COCO and ADE20K gains over CNNs, ViTs, MLPs, GNNs and Mambas.
Problem
Standard CNNs use fixed local windows (e.g., 7x7 kernels), which limits their flexibility and ability to capture long-range dependencies in images.
Approach
Treat every patch as a cluster center, pick its K-1 nearest patches by ℓ2 distance (computed on subsampled feature subvectors for efficiency), gather them into clusters and apply shared depthwise convolution over each cluster, so each patch can belong to multiple clusters and receive globally selected context.
Key results
- DCCNeXt-B2 reaches 82.8% ImageNet-1K top-1 with 26.7M params and 4.0G FLOPs, versus ConvNeXt V2-T at 82.5% with 28.6M params.
- On COCO Mask R-CNN 1x, DCCNeXt-B2 gets 44.6 APb vs 42.5 for ConvNeXt V2-T; on ADE20K Semantic FPN it reaches 44.5 mIoU vs 43.7.
- Subvector sampling cuts FLOPs at 1280x800 from 79.5G to 39.2G with comparable accuracy (78.7% on ImageNet for DCCNeXt-B1 variants).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DCCNeXt-B4 instead of QFormerh-B | ImageNet-1K classification · ImageNet-1K val | top-1 accuracy: 84.1 -> 84.5 | Helped |
| Use DCCNeXt-B2 instead of ConvNeXt V2-T | ImageNet-1K classification · ImageNet-1K val | top-1 accuracy: 82.5 -> 82.8 | Helped |
| Use DCCNeXt-B2 instead of ConvNeXt V2-T | RetinaNet 1× · COCO val2017 | AP: 41.7 -> 43.0 | Helped |
| Use DCCNeXt-B2 instead of ConvNeXt V2-T | Mask R-CNN 1× · COCO val2017 | APb: 42.5 -> 44.6 | Helped |
| Use DCCNeXt-B2 instead of ConvNeXt V2-T | Mask R-CNN 1× · COCO val2017 | APm: 38.8 -> 40.6 | Helped |
| Use DCCNeXt-B2 instead of ConvNeXt V2-T | Semantic FPN · ADE20K val | mIOU: 43.7 -> 44.5 | Helped |
| Add dynamic clustering convolution (ablation) | DCCNeXt-B1 · ImageNet-1K | top-1 accuracy: 75.3% -> 77.4% | Helped |
| Add convolution FFN (ablation) | DCCNeXt-B1 with DCConv · ImageNet-1K | top-1 accuracy: 77.4% -> 78.7% | Helped |
| Add subvector sampling (ablation) | DCCNeXt-B1 with ConvFFN · ImageNet-1K | top-1 accuracy: 78.7% -> 78.7% | No clear effect |
| Add subvector sampling (ablation) | DCCNeXt-B1 with ConvFFN · High-resolution input 1280 × 800 | FLOPs: 79.5 G -> 39.2 G | Helped |
| Set convolution kernel size K to 4 (ablation) | DCCNeXt-B1 · ImageNet-1K | top-1 accuracy: 78.7 -> 78.1 | Hurt |
| Set convolution kernel size K to 8 (ablation) | DCCNeXt-B1 · ImageNet-1K | top-1 accuracy: 78.7 -> 78.5 | Hurt |
Surprising findings
- Removing dynamic clustering convolution drops ImageNet top-1 only from 77.4% to 75.3% in the ablation, a modest gap for the claimed core module.
- Larger kernel sizes beyond 12 gave saturation rather than improvement, with best results (78.7%) at kernel sizes 12 and 18 in stages 3 and 4.
Limitations
- Global clustering still has quadratic cost in patch count, mitigated only by subvector sampling.
- Hardware and training compute are not reported in the text.
- Some table entries are inconsistent (e.g., DCCNeXt-B2 params listed as 33.8M for RetinaNet, 43.7M for Mask R-CNN).
Experimental setup
Datasets: ImageNet-1K COCO 2017 ADE20K
Benchmarks: ImageNet-1K COCO 2017 ADE20K
Metrics: top-1 accuracy AP APb APm mIoU FLOPs parameters
Baselines: ConvNeXt ConvNeXt V2 SLaK Swin Transformer Slide-PVT CycleMLP ViHGNN VMamba Context-Cluster FEC
Builds on: ConvNeXt Context-Cluster Mask R-CNN Semantic FPN
Closest papers by meaning
- Image as Set of Points (ICLR 2023 · similarity 0.86)
- Dynamic Convolution: Attention Over Convolution Kernels (CVPR 2020 · similarity 0.86)
- Joint Unsupervised Learning of Deep Representations and Image Clusters (CVPR 2016 · similarity 0.86)
- Unveiling the Unseen: Identifiable Clusters in Trained Depthwise Convolutional Kernels (ICLR 2024 · similarity 0.85)
- Neural Clustering based Visual Representation Learning (CVPR 2024 · similarity 0.83)
- From Logits to Hierarchies: Hierarchical Clustering made Simple (ICML 2025 · similarity 0.83)
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.