Reproduce “Unlocking the Potential of Vision Foundation Models for Multi-Task Dense Prediction via Mixture of Task Adapters” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yunzhi Zhuge, Xinzhuo Yu, Sitong Gong, Lu Zhang, Pingping Zhang, Huchuan Lu
D²MoE adapts frozen DINOv3 for multi-task dense prediction through shared low-rank updates and sparse task-conditioned experts, while combining adapted features with frozen-backbone features. On NYUD-v2, its ViT-L model improves aggregate performance by 11.34% over a frozen-backbone, heads-only baseline, and segmentation mIoU rises from 63.15 to 65.91. The results cover three benchmarks, not universal superiority or faster inference.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.