Related Experiment Video
Updated: Oct 10, 2025
Robotized Testing of Camera Positions to Determine Ideal Configuration for Stereo 3D Visualization of Open-Heart Surgery
Published on: August 12, 2021
Unsupervised learning of depth estimation from imperfect rectified stereo laparoscopic images
Huoling Luo1, Congcong Wang2, Xingguang Duan3
1Research Lab for Medical Imaging and Digital Surgery, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China; Shenzhen College of Advanced Technology, University of Chinese Academy of Sciences, Shenzhen, China.
This study introduces an unsupervised deep learning method for accurate depth estimation from stereo images, even with imperfect camera alignment. The approach effectively handles geometric inaccuracies, improving depth map generation for real-world applications.
Area of Science:
- Computer Vision
- Machine Learning
- 3D Reconstruction
Background:
- Learning-based depth estimation relies on precise stereo rectification.
- Inaccurate camera parameters degrade performance in unsupervised methods.
- A novel approach is needed to handle imperfect stereo alignment.
Purpose of the Study:
- To develop an unsupervised depth estimation method robust to imperfect stereo rectification.
- To enable accurate depth estimation without precise camera calibration.
- To improve the performance of learning-based depth estimation in real-world scenarios.
Main Methods:
- An unsupervised deep convolutional network is proposed for dense disparity map generation.
- A vertical correction module compensates for geometric misalignments.
- Generative adversarial networks (GANs) with a residual mask refine image reconstruction and depth accuracy.
Main Results:
- The model successfully estimates depth from imperfectly rectified stereo images.
- Validation on the SCARED dataset yielded an average Mean Absolute Error (MAE) of 3.054 mm.
- The method demonstrates robustness against geometric calibration errors.
Conclusions:
- The proposed model effectively performs depth estimation on stereo images with imperfect rectification.
- This work advances unsupervised learning for depth estimation by addressing geometric inaccuracies.
- The method offers a practical solution for depth sensing in scenarios with uncalibrated or imperfectly calibrated stereo cameras.
