Transformer-Based Approach for Monocular Depth Estimation Using DPT and DepthFormer

In computer vision tasks, depth estimation is essential because it allows systems to recognize and understand spatial relationships in an image. The study employs the DepthFormer model to generate depth maps from monocular images, effectively distinguishing between foreground and background objects based on their relative distances. The predicted depth map is found to accurately represent depth variations within the scene.With the smallest MAE (mean absolute error) score of 0.058, the DepthFormer model exhibited outstanding efficiency, which is significantly better compared to other depth estimation models. The real-time images depicting scenarios relevant to autonomous vehicle navigation, such as pedestrians crossing the road and vehicles moving on the street, were tested using the DepthFormer model. The model successfully generated accurate depth maps for these real-world scenes, demonstrating its capability to perceive depth effectively in dynamic environments. The results further validate the model's robustness and applicability in autonomous driving systems, where precise depth estimation is crucial for obstacle detection and navigation.