Employing Vision Transformer for Monocular Depth Estimation

Monocular depth estimation (MDE) is an important task in computer vision, it enables a range of applications like robotic navigation, augmented reality and also used in surgical guidance. This paper shows the use of ViTs for MDE and focuses on the Depth Anything V2(DA-V2) model. DA-V2 model leverages a teacher-student training approach and uses the synthetic and unlabeled real images which are designed to enhance the accuracy of depth estimation. DA-V2 is fine-tuned on EndoSLAM dataset and contains challenging visual scenes from endoscopic images of the stomach, colon and small intestine to check the adaptability of the model to medical depth estimation. Our experiment shows that the fine-tuned DA-V2 outperforms all the conventional models like Endo-SfMLearner and achieves reductions in root mean squared error(RMSE) and absolute relative error, across all anatomical datasets. DA-V2 achieves the RMSE values 1.245, 1.028, and 2.021 for the stomach, colon, and intestine datasets, respectively as compared to the SfMLearner’s 6.387, 6.657, and 14.787. The findings show that ViTs are capable of achieving higher accuracy and generalization in monocular depth estimation tasks. This study highlights the effectiveness of DA-V2 and Vision Transformers in Monocular Depth Estimation and opens up possibilities for further model optimization.