Research map / Detection and segmentation / CNN features
Recombinator Networks: Learning Coarse-To-Fine Feature Aggregation
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 final-only branch summation with coarse-to-fine feature concatenation | SumNet / facial keypoint localization · AFLW | normalized landmark mean error (%): 6.27 -> 5.60 | Helped |
| Replace final-only branch summation with coarse-to-fine feature concatenation | SumNet / facial keypoint localization · AFW | normalized landmark mean error (%): 6.33 -> 5.36 | Helped |
| Add denoising keypoint model (ablation) | RCN, 300W 49 keypoints · 300W Common, IBUG, and Fullset | normalized landmark mean error (%): 2.64 -> 2.59; 5.10 -> 4.81; 3.88 -> 3.76 | Helped |
| Add denoising keypoint model (ablation) | RCN, 300W 68 keypoints · 300W Common, IBUG, and Fullset | normalized landmark mean error (%): 4.70 -> 4.67; 9.00 -> 8.44; 5.54 -> 5.41 | Helped |
| Add training-time occlusion preprocessing (ablation) | SumNet (5 branch) · AFLW and AFW | normalized landmark mean error (%): 6.42 -> 6.29; 6.53 -> 6.34 | Helped |
| Add training-time occlusion preprocessing (ablation) | RCN (6 branch) · AFLW and AFW | normalized landmark mean error (%): 6.00 -> 5.60; 5.98 -> 5.36 | Helped |
| Increase RCN from 5 branches to 6 branches (ablation) | RCN · AFLW and AFW | normalized landmark mean error (%): 6.11 -> 6.00; 6.05 -> 5.98 | Helped |
| Increase RCN from 6 branches to 7 branches (ablation) | RCN · AFLW and AFW | normalized landmark mean error (%): 6.00 -> 6.17; 5.98 -> 6.12 | Hurt |
| Add skip connections from all coarser branches (ablation) | RCN (6 branch with occlusion) · AFLW and AFW | normalized landmark mean error (%): 5.60 -> 5.63; 5.36 -> 5.56 | Mixed |
| Use only the coarsest branch instead of all branches (ablation) | RCN · AFLW and AFW | normalized landmark mean error (%): 6.37 -> 10.61; 6.43 -> 10.89 | Hurt |
| Use only the finest branch instead of all branches (ablation) | RCN · AFLW and AFW | normalized landmark mean error (%): 6.37 -> 46.87; 6.43 -> 48.61 | Hurt |
| Use only the two coarsest branches instead of all branches (ablation) | RCN · AFLW and AFW | normalized landmark mean error (%): 6.37 -> 9.31; 6.43 -> 9.44 | Hurt |
Closest papers by meaning
- RNNPool: Efficient Non-linear Pooling for RAM Constrained Inference (NeurIPS 2020 · similarity 0.87)
- S3Pool: Pooling With Stochastic Spatial Sampling (CVPR 2017 · similarity 0.87)
- TI-Pooling: Transformation-Invariant Pooling for Feature Learning in Convolutional Neural Networks (CVPR 2016 · similarity 0.87)
- Beyond Skip Connection: Pooling and Unpooling Design for Elimination Singularities (AAAI 2025 · similarity 0.85)
- Implicit Bias of Mirror Flow in Homogeneous Neural Networks: Sparse and Dense Feature Learning (NeurIPS 2026 · similarity 0.85)
- SparX: A Sparse Cross-Layer Connection Mechanism for Hierarchical Vision Mamba and Transformer Networks (AAAI 2025 · similarity 0.84)
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.