Loading...

Supervised Monocular Depth Estimation using Deep Neural Networks

Ramezanian, Vida | 2025

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 58050 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Kasaei, Shohreh
  7. Abstract:
  8. In this project, we address monocular depth estimation, a fundamental challenge in understanding environmental scenes. Estimating depth from a single image can be particularly challenging due to out-of-distribution information such as occlusions, unexpected objects, and varying lighting conditions, all of which can disrupt model performance. Despite advancements in monocular depth estimation through the development of deep neural networks, challenges remain in accurately estimating depth maps and object boundaries. This research introduces a method aimed at improving the performance and reliability of monocular depth estimation models by leveraging epistemic uncertainty estimation in the training process. A baseline model based on a convolutional encoder-decoder architecture, which typically estimates depth maps, is considered. To enhance this model, multiple outputs are generated using a Monte Carlo sampling approach, and the variance of these outputs is calculated to derive epistemic uncertainty maps. These maps indicate the model's confidence in its predictions across different regions of the image. By focusing on unknown or out-of-distribution data, they enable improved generalizability of the model. Subsequently, these uncertainty maps are integrated into the loss function as a weighting mask, allowing pixel weights to be adjusted in areas with high uncertainty. This encourages the model to concentrate more on improving these regions. Consequently, the training process benefits from a feedback mechanism based on uncertainty, enabling the model to adaptively adjust pixel weighting in high-uncertainty regions and thereby enhance accuracy and flexibility at each training stage. The results demonstrate significant improvements in quantitative accuracy metrics, including a 6% improvement in the regression-based baseline model and a 2% improvement in the classification-based baseline model, as measured by the relative squared error metric. Furthermore, the proposed method, due to its reliance on general uncertainty concepts and straightforward integration into the loss function, can be easily adapted with minor modifications to other encoder-decoder-based monocular depth estimation networks
  9. Keywords:
  10. Deep Neural Networks ; Uncertainty Estimation ; Supervised Monocular Depth Estimation ; Uncertainty-Aware Monocular Depth Estimation

 Digital Object List

 Bookmark

...see more