Unitary Camera Proximal Intensity Estimation using DEVIANT

Accurate depth estimation is crucial for applications like autonomous driving and augmented reality. This project proposes a novel approach to enhance monocular depth estimation (MDE) by integrating annotated and unannotated images. The methodology involves training a student model on a dataset combining labeled and pseudo-labeled data, leveraging a teacher model's knowledge trained on an extensive dataset. Data acquisition focuses on high-quality annotated datasets like KITTI and NuScenes. Preprocessing involves image resizing and normalization, with data augmentation techniques for diversity. Network architecture includes CNNs and ViTs for feature extraction, with FPNs for multi-scale feature representation. The project utilizes the DEVIANT model with the vanilla architecture for enhanced results, overcoming limitations of existing MiDAS architectures. Future scope includes applying the methodology to other vision tasks and incorporating more sophisticated network architectures. The propsed solution overcomes teh use of LIDAR and multi camera setups for such applications. Consequently, the proposed solution demonstrates outstanding zero-shot depth estimation capability, and also provides a promising foundation for subsequent tasks involving metric depth estimation and semantic segmentation.