Research map / Architectures and efficiency / Mixture of experts
Learning Robust Stereo Matching in the Wild with Selective Mixture-of-Experts
Yun Wang, Longguang Wang, Chenghao Zhang, Yongjian Zhang, Zhanjie Zhang, Ao Ma, Chenyou Fan, Tin Lun Lam, Junjie Hu
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 |
|---|---|---|---|
| add MoE LoRA and MoE Adapter to the frozen VFM (ablation) | SMoEStereo (DAMV2 ViT-base) · KITTI 2015 (RVC ablation) | D1_All: 2.01 -> 1.57 | Improved |
| add MoE LoRA alone (ablation) | SMoEStereo (DAMV2 ViT-base) · Middlebury (RVC ablation) | Bad 2.0: 7.67 -> 4.80 | Improved |
| add MoE Adapter alone (ablation) | SMoEStereo (DAMV2 ViT-base) · Middlebury (RVC ablation) | Bad 2.0: 7.67 -> 4.68 | Improved |
| add the learned decision network to prune MoE modules (ablation) | SMoEStereo (DAMV2 ViT-base) · KITTI 2015 (RVC ablation) | D1_All and extra params: 1.57 -> 1.51; params 6.77 -> 2.86 | Improved |
| replace learned usage policy with a random policy of similar cost (ablation) | SMoEStereo (DAMV2 ViT-base) · KITTI 2015 (RVC ablation) | D1_All: 1.51 -> 1.69 | Hurt |
| use mixed-rank MoE LoRA experts instead of homogeneous fixed rank 4 (ablation) | SMoE MoE LoRA · KITTI 2015 (RVC ablation) | D1_All: 1.77 -> 1.51 | Improved |
| use sparse expert selection instead of aggregating all experts (Multi-E) (ablation) | SMoEStereo · Middlebury (RVC ablation) | Bad 2.0: 4.37 -> 4.12 | Improved |
| use sparse expert selection instead of Multi-E on KITTI 2015 (ablation) | SMoEStereo · KITTI 2015 (RVC ablation) | D1_All: 1.46 -> 1.51 | Hurt |
| selective activation instead of all experts for speed (ablation) | SMoEStereo · KITTI inference speed | inference speed: 4.68 iter/s -> 5.26 iter/s | Improved |
| use SMoE instead of LoRA for adapting DAMV2 | DAMV2 ViT-base · Middlebury zero-shot | Bad 2.0: 7.67 -> 7.05 | Improved |
| use PEFT (LoRA) instead of full fine-tuning | DAMV2 ViT-base · KITTI 2015 zero-shot | Bad 3.0: 14.3 -> 5.03 | Improved |
| scale VFM capacity from ViT-Small to ViT-Large | SMoEStereo (DAMV2) · ETH3D zero-shot | Bad 1.0: 2.43 -> 1.87 | Improved |
Closest papers by meaning
- All-in-One: Transferring Vision Foundation Models into Stereo Matching (AAAI 2025 · similarity 0.91)
- AdaStereo: An Efficient Domain-Adaptive Stereo Matching Approach (IJCV 2022 · similarity 0.90)
- Stereo Matching Using Multi-Level Cost Volume and Multi-Scale Feature Constancy (TPAMI 2019 · similarity 0.90)
- Learning Representations from Foundation Models for Domain Generalized Stereo Matching (ECCV 2024 · similarity 0.90)
- Hierarchical Object-Aware Dual-Level Contrastive Learning for Domain Generalized Stereo Matching (NeurIPS 2024 · similarity 0.89)
- Stereo Risk: A Continuous Modeling Approach to Stereo Matching (ICML 2024 · similarity 0.89)
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.
