The "Roaring 20s" of visual recognition began with the introduction of Vision Transformers (ViTs), which quickly superseded ConvNets as the state-of-the-art image classification model. A vanilla ViT, on the other hand, faces difficulties when applied to general computer vision tasks such as object detection and semantic segmentation. It is the hierarchical Transformers (e.g., Swin Transformers) that reintroduced several ConvNet priors, making Transformers practically viable as a generic vision backbone and demonstrating remarkable performance on a wide variety of vision tasks. However, the effectiveness of such hybrid approaches is still largely credited to the intrinsic superiority of Transformers, rather than the inherent inductive biases of convolutions. In this work, we reexamine the design spaces and test the limits of what a pure ConvNet can achieve. We gradually "modernize" a standard ResNet toward the design of a vision Transformer, and discover several key components that contribute to the performance difference along the way. The outcome of this exploration is a family of pure ConvNet models dubbed ConvNeXt. Constructed entirely from standard ConvNet modules, ConvNeXts compete favorably with Transformers in terms of accuracy and scalability, achieving 87.8% ImageNet top-1 accuracy and outperforming Swin Transformers on COCO detection and ADE20K segmentation, while maintaining the simplicity and efficiency of standard ConvNets.
核心贡献 · Key contributions
提出 ConvNeXt，一个纯卷积网络系列，在准确性和可扩展性上与 Transformer 竞争。 Proposes ConvNeXt, a pure ConvNet family that competes favorably with Transformers in accuracy and scalability.
系统地将标准 ResNet 现代化为视觉 Transformer 设计，识别关键性能组件。 Systematically modernizes a standard ResNet towards vision Transformer design, identifying key performance components.
在 ImageNet 上达到 87.8% top-1 准确率，并在 COCO 检测和 ADE20K 分割上超越 Swin Transformer。 Achieves 87.8% ImageNet top-1 accuracy and outperforms Swin Transformers on COCO detection and ADE20K segmentation.
证明纯卷积网络在多个基准测试上可与分层视觉 Transformer 匹敌。 Demonstrates that pure ConvNets can match hierarchical vision Transformers across multiple benchmarks.
显示 ConvNeXt 在保持标准卷积网络简单性和效率的同时达到最先进性能。 Shows ConvNeXt maintains simplicity and efficiency of standard ConvNets while achieving state-of-the-art performance.
提供关于大核尺寸、倒瓶颈和归一化等设计选择的详细消融研究。 Provides detailed ablation study on design choices like large kernel size, inverted bottleneck, and normalization.
局限 · Limitations
ConvNeXt 可能不太适合需要跨模态交叉注意力的多模态学习。 ConvNeXt may be less suited for multi-modal learning requiring cross-attention across modalities.
Transformer 对于需要离散化、稀疏或结构化输出的任务可能更灵活。 Transformers may be more flexible for tasks requiring discretized, sparse, or structured outputs.
大型模型变体需要大量算力资源，增加碳排放。 Large model variants require significant compute resources, increasing carbon emissions.
ConvNeXt 与 Transformer 的鲁棒性和公平性需要进一步研究。 Robustness and fairness of ConvNeXt vs. Transformers require further investigation.
ConvNeXt 受益于大规模预训练，可能引发数据偏差问题。 ConvNeXt benefits from large-scale pre-training, which may raise data bias concerns.
论文章节 · Sections（共 26）
摘要Abstract
1 引言1 Introduction
2 现代化卷积网络：路线图2 Modernizing a ConvNet: a Roadmap
2.1 训练技巧2.1 Training Techniques
2.2 宏观设计2.2 Macro Design
2.3 ResNeXt 化2.3 ResNeXt-ify
2.4 倒置瓶颈2.4 Inverted Bottleneck
2.5 大卷积核2.5 Large Kernel Sizes
2.6 微观设计2.6 Micro Design
3 ImageNet 上的实证评估3 Empirical Evaluations on ImageNet
3.1 设置3.1 Settings
3.2 结果3.2 Results
3.3 各向同性 ConvNeXt 与 ViT 对比3.3 Isotropic ConvNeXt vs. ViT
4 下游任务实证评估4 Empirical Evaluation on Downstream Tasks
5 相关工作5 Related Work
6 结论6 Conclusions
附录Appendix
A.1 ImageNet（预）训练A.1 ImageNet (Pre-)training
A.2 ImageNet 微调A.2 ImageNet Fine-tuning
A.3 下游任务A.3 Downstream Tasks
附录 B 鲁棒性评估Appendix B Robustness Evaluation
附录 C 现代化 ResNet：详细结果Appendix C Modernizing ResNets: detailed results
附录 D 详细架构Appendix D Detailed Architectures
附录 E A100 GPU 基准测试Appendix E Benchmarking on A100 GPUs