Research map / Architectures and efficiency / Pruning and efficient tuning
Unlocking the Potential of Vision Foundation Models for Multi-Task Dense Prediction via Mixture of Task Adapters
Yunzhi Zhuge, Xinzhuo Yu, Sitong Gong, Lu Zhang, Pingping Zhang, Huchuan Lu
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 MoTA (task adapters) to frozen backbone (ablation) | MTL baseline · PASCAL-Context (frozen DINOv3 ViT-L, ablation) | Delta_m (%): 0.00 -> 6.10 | Improved |
| add SPI (semantic prior injector) alone (ablation) | MTL baseline · PASCAL-Context (frozen DINOv3 ViT-L, ablation) | Delta_m (%): 0.00 -> 5.66 | Improved |
| add SPI on top of MoTA (ablation) | MoTA · PASCAL-Context (frozen DINOv3 ViT-L, ablation) | Delta_m (%): 6.10 -> 8.88 | Improved |
| add TG-LoRA to MoTA (ablation) | MoTA w/o TG-LoRA · PASCAL-Context (frozen DINOv3 ViT-L, ablation) | Delta_m (%): 7.74 -> 8.88 | Improved |
| add TA-MoE to MoTA (ablation) | MoTA w/o TA-MoE · PASCAL-Context (frozen DINOv3 ViT-L, ablation) | Delta_m (%): 7.97 -> 8.88 | Improved |
| insert MoE at QKV and projection instead of none (ablation) | TA-MoE · PASCAL-Context (frozen DINOv3 ViT-L, ablation) | Delta_m (%): 7.97 -> 8.88 | Improved |
| also insert MoE in FFN beyond QKV+Proj (ablation) | TA-MoE · PASCAL-Context (frozen DINOv3 ViT-L, ablation) | Delta_m (%): 8.88 -> 8.68 | Hurt |
| place experts in backbone instead of decoder (ablation) | D2MoE vs Decoder MoE · PASCAL-Context (frozen DINOv3 ViT-L, matched budget) | Delta_m (%): 5.90 -> 8.88 | Improved |
| task-aware routing vs shared LoRA at matched budget (ablation) | D2MoE · PASCAL-Context (frozen DINOv3 ViT-L, matched budget) | Delta_m (%): 8.48 -> 8.88 | Improved |
| increase TG-LoRA rank from 2 to 16 (ablation) | TG-LoRA · PASCAL-Context (frozen DINOv3 ViT-L, ablation) | Delta_m (%): 8.07 -> 8.88 | Improved |
| increase number of experts from 2 to 16, then 32 (ablation) | TA-MoE · PASCAL-Context (frozen DINOv3 ViT-L, ablation) | Delta_m (%): peak 8.88 at 16 experts; parsing 78.05 -> 77.64 at 32 | Mixed |
| scale backbone ViT-L to ViT-H (ablation) | D2MoE · PASCAL-Context (frozen DINOv3 ViT-L, ablation) | Saliency maxF: 85.55 -> 83.17 | Hurt |
Closest papers by meaning
- Parameter-, Memory-, Time-Efficient Multi-Task Dense Vision Adaptation (AAAI 2026 · similarity 0.90)
- VMT-Adapter: Parameter-Efficient Transfer Learning for Multi-Task Dense Scene Understanding (AAAI 2024 · similarity 0.89)
- Swiss Army Knife: Synergizing Biases in Knowledge from Vision Foundation Models for Multi-Task Learning (ICLR 2025 · similarity 0.89)
- Going Beyond Multi-Task Dense Prediction with Synergy Embedding Models (CVPR 2024 · similarity 0.87)
- TADFormer: Task-Adaptive Dynamic TransFormer for Efficient Multi-Task Learning (CVPR 2025 · similarity 0.87)
- Task-Conditioned Adaptation of Visual Features in Multi-Task Policy Learning (CVPR 2024 · similarity 0.86)
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.
