Research map / Architectures and efficiency / CNN design
ConvNeXt V2: Co-Designing and Scaling ConvNets With Masked Autoencoders
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use sparse convolution in the FCMAE encoder to block masked-region information leakage (ablation) | ConvNeXt-Base FCMAE · ImageNet-1K val | fine-tuning accuracy: 79.3 -> 83.7 | Helped |
| Replace the single ConvNeXt block decoder with a transformer decoder (ablation) | ConvNeXt-Base FCMAE · ImageNet-1K | pre-training time: 7.7 -> 8.5 hours | Hurt |
| Use a single ConvNeXt block rather than deeper decoder stacks (ablation) | ConvNeXt-Base FCMAE · ImageNet-1K val | fine-tuning accuracy: 1 block: 83.7; 4 blocks: 83.7; 12 blocks: 83.3 | Mixed |
| Use L2-norm aggregation instead of global average pooling in GRN (ablation) | ConvNeXt-Base · ImageNet-1K val | fine-tuning accuracy: 83.7 -> 84.6 | Helped |
| Keep the residual connection in GRN (ablation) | ConvNeXt-Base · ImageNet-1K val | fine-tuning accuracy: 84.0 -> 84.6 | Helped |
| Use GRN rather than LRN, BN, or LN (ablation) | ConvNeXt-Base FCMAE · ImageNet-1K val | fine-tuning accuracy: GRN 84.6; LRN 83.2; BN 80.5; LN 83.8 | Helped |
| Use GRN rather than SE or CBAM feature gating (ablation) | ConvNeXt-Base FCMAE · ImageNet-1K val | fine-tuning accuracy: SE 84.4; CBAM 84.5; GRN 84.6 | Helped |
| Use GRN during both pre-training and fine-tuning (ablation) | ConvNeXt-Base FCMAE · ImageNet-1K val | fine-tuning accuracy: 78.8 / 80.6 -> 84.6 | Helped |
| Combine global aggregation and normalization in GRN rather than using either component alone (ablation) | ConvNeXt V2-Base · ImageNet-1K val | fine-tuning accuracy: 83.9 -> 84.6 | Helped |
| Co-design the GRN architecture with FCMAE instead of applying FCMAE to ConvNeXt V1 | ConvNeXt-Base · ImageNet-1K val | top-1 accuracy: 83.7 -> 84.6 | Helped |
| Use GRN-enhanced ConvNeXt V2 with FCMAE rather than supervised ConvNeXt V1 | ConvNeXt V2-B · COCO val2017 | box AP: 50.3 -> 52.9 | Helped |
| Use GRN-enhanced ConvNeXt V2 with FCMAE rather than supervised ConvNeXt V1 | ConvNeXt V2-B · ADE20K | mIoU: 49.9 -> 52.1 | Helped |
Closest papers by meaning
- VideoMAC: Video Masked Autoencoders Meet ConvNets (CVPR 2024 · similarity 0.88)
- The Dynamic Duo of Collaborative Masking and Target for Advanced Masked Autoencoder Learning (AAAI 2025 · similarity 0.88)
- MogaNet: Multi-order Gated Aggregation Network (ICLR 2024 · similarity 0.87)
- Exploring Target Representations for Masked Autoencoders (ICLR 2024 · similarity 0.87)
- DECO: Unleashing the Potential of ConvNets for Query-based Detection and Segmentation (ICLR 2025 · similarity 0.87)
- UniRepLKNet: A Universal Perception Large-Kernel ConvNet for Audio Video Point Cloud Time-Series and Image Recognition (CVPR 2024 · similarity 0.86)
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.