Research map / Detection and segmentation / CNN features
Involution: Inverting the Inherence of Convolution for Visual Recognition
Check 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 |
|---|---|---|---|
| Replace ResNet-50 stem and trunk convolutions with involution | ImageNet classification · ImageNet val | Top-1 accuracy: 76.8 -> 78.4 | Helped |
| Replace the ResNet-50 backbone with RedNet-50 | RetinaNet · COCO val2017 | bbox AP: 36.6 -> 38.3 | Helped |
| Use involution instead of convolution in the neck with RedNet-50 backbone (ablation) | Faster R-CNN · COCO val2017 | bbox AP: 39.5 -> 40.2 | Helped |
| Replace convolution in the Faster R-CNN head with involution (ablation) | Faster R-CNN with involution backbone and neck · COCO val2017 | bbox AP: 40.2 -> 39.2 | Hurt |
| Use involution instead of convolution in the Mask R-CNN neck with RedNet-50 backbone (ablation) | Mask R-CNN · COCO val2017 | bbox AP: 40.2 -> 40.8 | Helped |
| Replace convolution in the Mask R-CNN head with involution (ablation) | Mask R-CNN with involution backbone and neck · COCO val2017 | bbox AP: 40.8 -> 39.6 | Hurt |
| Replace the ResNet-50 backbone with RedNet-50 | Semantic FPN · Cityscapes val | mean IoU: 74.5 -> 78.3 | Helped |
| Use involution instead of convolution in the Semantic FPN neck (ablation) | Semantic FPN with RedNet-50 · Cityscapes val | mean IoU: 78.3 -> 79.2 | Helped |
| Increase involution kernel size from 3 × 3 to 5 × 5 (ablation) | RedNet-50 · ImageNet val | Top-1 accuracy: 76.9 -> 77.4 | Helped |
| Increase involution kernel size from 5 × 5 to 7 × 7 (ablation) | RedNet-50 · ImageNet val | Top-1 accuracy: 77.4 -> 77.7 | Helped |
| Increase involution kernel size from 7 × 7 to 9 × 9 (ablation) | RedNet-50 · ImageNet val | Top-1 accuracy: 77.7 -> 77.8 | Helped |
| Share each involution kernel across 16 channels rather than using one channel per group (ablation) | RedNet-50 · ImageNet val | Top-1 accuracy: 77.9 -> 77.7 | No clear effect |
Closest papers by meaning
- Contextual Convolutional Networks (ICLR 2023 · similarity 0.91)
- Deep Residual Learning for Image Recognition (CVPR 2016 · similarity 0.90)
- Active Convolution: Learning the Shape of Convolution for Image Classification (CVPR 2017 · similarity 0.90)
- SparseFormer: Sparse Visual Recognition via Limited Latent Tokens (ICLR 2024 · similarity 0.88)
- KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 2024 · similarity 0.88)
- AsCAN: Asymmetric Convolution-Attention Networks for Efficient Recognition and Generation (NeurIPS 2024 · similarity 0.88)
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.