Research map / Vision-language and multimodal / Cross-modal learning
Deep Multimodal Fusion by Channel Exchanging
Yikai Wang, Wenbing Huang, Fuchun Sun, Tingyang Xu, Yu Rong, Junzhou Huang
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 CEN rather than the strongest reported baseline | Semantic segmentation · NYUDv2 | Mean IoU: 50.1 -> 52.5 | Improved |
| Use CEN rather than the strongest reported baseline | Semantic segmentation · SUN RGB-D | Mean IoU: 50.7 -> 51.1 | Improved |
| Use CEN fusion rather than all-layer self-attention | U-Net-256 · Shade+Texture → RGB | FID: 73.87 -> 62.63 | Improved |
| Share BN layers across modalities in addition to convolutions (ablation) | CEN · NYUDv2 | Ensemble Mean IoU: 48.0 -> 45.2 | Hurt |
| Share convolutional parameters while keeping separate BNs (ablation) | Unshared RGB-D subnetworks · NYUDv2 | Ensemble Mean IoU: 47.6 -> 48.0 | Improved |
| Add directed half-channel exchange with ℓ1 regulation (ablation) | Shared-convolution CEN without exchange · NYUDv2 | Ensemble Mean IoU: 47.7 -> 51.1 | Improved |
| Allow exchange on all channels rather than half (ablation) | CEN · NYUDv2 | Ensemble Mean IoU: 51.1 -> 49.8 | Hurt |
| Use fixed 30% exchange without ℓ1 regulation and threshold (ablation) | CEN · NYUDv2 | Mean IoU: 47.2 | Hurt |
| Exchange only at down-sampling stages rather than throughout the encoder (ablation) | CEN · NYUDv2 | Mean IoU: 48.6 vs 51.1 | Hurt |
| Replace exchange with direct activation summation (ablation) | CEN · NYUDv2 | Mean IoU: 48.1 | Hurt |
| Replace expert ensemble with a concat-fusion block (ablation) | CEN · NYUDv2 | Mean IoU: 51.1 -> 50.8 | Hurt |
| Exchange channels randomly rather than using learned importance (ablation) | CEN · NYUDv2 | Mean IoU: 46.8 | Hurt |
Closest papers by meaning
- Channel Exchanging Networks for Multimodal and Multitask Dense Image Prediction (TPAMI 2022 · similarity 0.93)
- GeminiFusion: Efficient Pixel-wise Multimodal Fusion for Vision Transformer (ICML 2024 · similarity 0.87)
- Keep the Balance: A Parameter-Efficient Symmetrical Framework for RGB+X Semantic Segmentation (CVPR 2025 · similarity 0.87)
- Understanding Unimodal Bias in Multimodal Deep Linear Networks (ICML 2024 · similarity 0.86)
- Beyond Modality Fusion: Deep Ensembles for Multimodal Classification (NeurIPS 2026 · similarity 0.85)
- Embracing Unimodal Aleatoric Uncertainty for Robust Multimodal Fusion (CVPR 2024 · similarity 0.85)
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.
