Reproduce “Empowering DINO Representations for Underwater Instance Segmentation via Aligner and Prompter” 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
Zhiyang Chen, Chen Zhang, Hao Fang, Runmin Cong
DiveSeg introduces a framework that adapts DINOv2 for underwater instance segmentation, utilizing an AquaStyle Aligner for scene-level domain adaptation and an ObjectPrior Prompter for object-level prior integration. The framework achieves state-of-the-art performance on UIIS and USIS10K datasets, improving mAP on UIIS by 21.1% over previous methods while being more parameter efficient.
Key insights
- Decomposing VFM adaptation into distinct scene-level and object-level modules allows for targeted and effective domain transfer to challenging environments like underwater.
- Fourier decomposition of the image amplitude component effectively captures low-level underwater style characteristics, which can be injected into the VFM backbone via cross-attention for efficient domain adaptation.
- Dynamically generated pseudo masks can serve as powerful object-level priors, enhancing instance discrimination for objects underrepresented in general pre-training data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.