AURA-Depth: Attention-Based Uncertainty Reduction and Feature Aggregation Depth Network

Reliable depth information is crucial in autonomous driving technology. However, monocular depth estimation often faces inherent limitations such as scale ambiguity and lack of depth cues. To address these issues, we propose a novel approach that fuses two methods: triangulated depth estimation derived from the optical flow of consecutive frames, and dense depth estimation from a single image. In the initial task, we use a dense depth network to extract a depth map from the current frame and calculate a triangulated depth map through the optical flow and relative pose differences between previous frames. While triangulated depth estimation may lack detail in object-specific regions, it provides high accuracy for the overall scene depth. Conversely, dense depth estimation excels at extracting the relative depth of objects, making these methods complementary. The main task employs a multi-branch architecture to process RGB image, dense depth map, and triangulated depth map. The output of each encoder branch’s residual module is fused using a Multimodal Cross Attention (MCA) module, guided by the uncertainty calculated from the absolute difference between dense depth features and triangulated depth features. This process effectively integrates the features of each branch. The final depth map is generated by a decoder that processes the fused features. Our approach significantly improves depth estimation accuracy and reliability by leveraging the complementary strengths of uncertainty-guided dense depth maps and triangulated depth maps, overcoming the limitations of traditional monocular methods. This method demonstrates substantial improvements in various computer vision tasks, providing a robust solution.