EndoMINI: Self-Supervised Depth Estimation for Endoscopic Scenes
A novel self-supervised framework called EndoMINI has been proposed to address challenges in depth estimation for endoscopic surgeries. The method, detailed in a paper on arXiv (2608.00415), introduces a mixture of low-rank experts (MiLoRE) for parameter-efficient fine-tuning, enhancing model adaptation to diverse endoscopic scenes. Additionally, an intrinsic image alignment (IIA) technique, supported by a novel intrinsic image decomposition network, is integrated into the training loss to mitigate the effects of light reflectance. The approach was evaluated on the SCARED dataset for supervised depth estimation and on Hamlyn and SERV-CT datasets for zero-shot depth estimation, demonstrating state-of-the-art performance. This research contributes to improving 3D perception in endoscopic procedures, potentially aiding surgical navigation and automation.
Key facts
- EndoMINI is a self-supervised framework for depth estimation in endoscopic scenes.
- MiLoRE (mixture of low-rank experts) enables parameter-efficient fine-tuning.
- IIA (intrinsic image alignment) is introduced to reduce light reflectance interference.
- A novel intrinsic image decomposition network supports the IIA technique.
- Evaluation was conducted on SCARED, Hamlyn, and SERV-CT datasets.
- The method achieves state-of-the-art results in supervised and zero-shot depth estimation.
- The paper is available on arXiv with ID 2608.00415.
- The research targets challenges like illumination interference and feature diversity in endoscopy.
Entities
Institutions
- arXiv