Research map / Detection and segmentation / CNN features
Reversible Column Networks
Yuxuan Cai, Yizhuang Zhou, Qi Han, Jianjian Sun, Xiangwen Kong, Jun Li, Xiangyu Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 reversible multi-column connections instead of non-reversible HRNet-style fusion (ablation) | RevCol-S · ImageNet-1K | Top-1 accuracy: 78.8 -> 82.2 | Improved |
| Add multi-column fusion to a single-column reversible network (ablation) | RevCol-S · ImageNet-1K | Top-1 accuracy: 81.6 -> 82.2 | Improved |
| Add intermediate supervision (ablation) | RevCol-T · ImageNet-1K | Top-1 accuracy: 81.4 -> 82.2 (+0.8) | Improved |
| Add intermediate supervision (ablation) | RevCol-S · ImageNet-1K | Top-1 accuracy: 83.0 -> 83.5 (+0.5) | Improved |
| Add intermediate supervision (ablation) | RevCol-B · ImageNet-1K | Top-1 accuracy: 83.2 -> 84.1 (+0.9) | Improved |
| Add intermediate supervision (ablation) | RevCol-T · COCO, 1× schedule | APbox: 48.3 -> 48.8 (+0.6) | Improved |
| Add intermediate supervision (ablation) | RevCol-S · COCO, 1× schedule | APmask: 43.8 -> 43.8 (+0.0) | No clear effect |
| Increase convolution kernel from 3×3 to 5×5 (ablation) | RevCol-T · ImageNet-1K | Top-1 accuracy: 82.2 -> 82.5 | Improved |
| Increase convolution kernel from 3×3 to 5×5 (ablation) | RevCol-T · COCO, 1× schedule | APbox: 48.8 -> 49.5 | Improved |
| Increase convolution kernel from 5×5 to 11×11 (ablation) | RevCol-T · ImageNet-1K | Top-1 accuracy: 82.5 -> 82.5 | No clear effect |
| Scale the model to 20 columns under a similar compute budget (ablation) | RevCol · ImageNet-1K | Top-1 accuracy: 82.2 -> 81.0 | Hurt |
| Use a four-column reversible design rather than a single-column model (ablation) | RevCol · ImageNet-1K | Top-1 accuracy: 81.9 -> 82.2 | Improved |
Closest papers by meaning
- RevColV2: Exploring Disentangled Representations in Masked Image Modeling (NeurIPS 2023 · similarity 0.88)
- The Reversible Residual Network: Backpropagation Without Storing Activations (NeurIPS 2017 · similarity 0.86)
- Dr2Net: Dynamic Reversible Dual-Residual Networks for Memory-Efficient Finetuning (CVPR 2024 · similarity 0.82)
- Reversible Decoupling Network for Single Image Reflection Removal (CVPR 2025 · similarity 0.82)
- RUN: Reversible Unfolding Network for Concealed Object Segmentation (ICML 2025 · similarity 0.81)
- PETRA: Parallel End-to-end Training with Reversible Architectures (ICLR 2025 · similarity 0.80)
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.
