Reproduce “EndoDAC: Efficient Adapting Foundation Model for Self-Supervised Depth Estimation from Any Endoscopic Camera” 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
Beilei Cui, Mobarakol Islam, Long Bai, An Wang, Hongliang Ren
EndoDAC introduces an efficient, self-supervised framework that adapts foundation models for monocular depth estimation in endoscopic surgery, operating effectively even with unknown camera intrinsic parameters. The method achieves state-of-the-art depth accuracy on SCARED and Hamlyn datasets using only 1.6% of the foundation model's parameters.
Key insights
- Foundation models can be efficiently adapted for self-supervised regression tasks in specialized medical domains with remarkably few trainable parameters (e.g., 1.6% of total model parameters).
- The combination of Dynamic Vector-Based LoRA (DV-LoRA) and Convolutional Neck blocks effectively bridges the domain gap for ViT-based foundation models in endoscopy while preserving high-frequency details.
- Simultaneous self-supervised estimation of depth, ego-motion, and camera intrinsic parameters from monocular video is feasible and yields accurate results, eliminating the need for prior camera calibration.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.