Research map / Label-efficient and robust learning / Recognition and long-tail
MTI-Net: Multi-Scale Task Interaction Networks for Multi-Task Learning
Simon Vandenhende, Stamatios Georgoulis, Luc Van Gool
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 feature propagation module (FPM) (ablation) | MTI-Net · NYUD-v2 (HRNet-18, seg+depth) | multi-task gain Δm (%): +3.85% -> +6.40% | Improved |
| add FPM with normals and edges auxiliary tasks (ablation) | MTI-Net · NYUD-v2 (HRNet-18, seg+depth) | multi-task gain Δm (%): +4.48% -> +10.91% | Improved |
| replace single-scale PAD-Net distillation with multi-scale distillation (no FPM) | PAD-Net · NYUD-v2 (HRNet-18, seg+depth) | multi-task gain Δm (%): -0.02% -> +3.85% | Improved |
| add backbone scales 1/8, 1/16 and 1/32 to the single 1/4 scale (ablation) | MTI-Net · NYUD-v2 number of scales (HRNet-18) | multi-task gain Δm (%): -0.02 -> +3.80 -> +5.53 -> +6.40 | Improved |
| add normals and edges auxiliary tasks to MTI-Net with FPM (ablation) | MTI-Net · NYUD-v2 (HRNet-18, seg+depth) | multi-task gain Δm (%): +6.40% -> +10.91% | Improved |
| add edge detection auxiliary task (ablation) | MTI-Net (s) · PASCAL (HRNet-18) | multi-task gain Δm (%): +3.35 -> +3.98 | Improved |
| add surface normals auxiliary task (ablation) | MTI-Net (s) · PASCAL (HRNet-18) | multi-task gain Δm (%): +3.35 -> +2.69 | Hurt |
| add surface normals auxiliary task (ablation) | MTI-Net (s) · PASCAL (R18-FPN) | multi-task gain Δm (%): +2.77 -> 2.52 | Hurt |
| use shared-encoder multi-task learning instead of single-task models | single-task baseline · NYUD-v2 (HRNet-18, seg+depth) | multi-task gain Δm (%): -1.71 | Hurt |
| train all five tasks with a shared encoder (MTL) instead of single-task models | single-task baseline · PASCAL (HRNet-18) | multi-task gain Δm (%): -3.70 | Hurt |
| replace ASTMT with MTI-Net on the same R50-FPN backbone | ASTMT · PASCAL (R50-FPN) | multi-task gain Δm (%): -0.87% -> +1.36% | Improved |
| replace PAD-Net with MTI-Net on PASCAL | PAD-Net · PASCAL (HRNet-18) | multi-task gain Δm (%): -3.08% -> +2.74% | Improved |
Later papers that compared with this method
1 later accepted paper reports MTI-Net as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Universal Representations: A Unified Look at Multiple Task and Domain Learning (IJCV 2023 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Towards Principled Task Grouping for Multi-Task Learning (NeurIPS 2026 · similarity 0.88)
- Towards Consistent Multi-Task Learning: Unlocking the Potential of Task-Specific Parameters (CVPR 2025 · similarity 0.88)
- Distribution Matching for Multi-Task Learning of Classification Tasks: A Large-Scale Study on Faces & Beyond (AAAI 2024 · similarity 0.87)
- Selective Task Group Updates for Multi-Task Optimization (ICLR 2025 · similarity 0.87)
- Quantifying Task Priority for Multi-Task Optimization (CVPR 2024 · similarity 0.87)
- Identifying and Mitigating Spurious Correlation in Multi-Task Learning (CVPR 2025 · similarity 0.87)
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.
