Introduction
Self-supervised monocular depth estimation is a critical task for autonomous systems, enabling robots and vehicles to perceive 3D structures from standard video without the need for expensive LiDAR labels. Traditional frameworks, such as Monodepth2, rely on warping consecutive video frames based on estimated depth and camera motion. However, these methods often struggle with the "static scene assumption," leading to blurry artifacts and depth discontinuities around moving objects.
This paper introduces a novel joint learning framework that treats moving objects not as noise to be masked, but as sources of "implicit cues." By leveraging temporal "unit streams" between frames, the proposed method achieves sharper, more accurate depth maps.
Proposed Methodology
The framework centers on two primary innovations: the Implicit Depth Cues Extractor (IDCE) and the High-dimensional Attention Module (HAM).
Implicit Depth Cues Extractor (IDCE)
In standard architectures, the depth and pose networks are largely isolated. The IDCE acts as a bridge, feeding shallow features from the PoseNet encoder—which inherently captures motion—into the DepthNet decoder. This allows the model to utilize motion cues as a "dynamic supplement," helping the network better define the contours of moving objects that might otherwise be blurred by the photometric reprojection process.
High-dimensional Attention Module (HAM)
To address uncertainty caused by occlusions and complex lighting, the HAM employs a Gaussian kernel to map local features into a high-dimensional space. This mechanism helps the model identify global correlations across the image, effectively suppressing noise and leading to a more robust estimation of the camera's ego-motion.
Experimental Evaluation
The model was trained on the KITTI dataset and CityScapes for urban environment recognition. To test its generalization capabilities, it was evaluated on the Make3D dataset, which consists of environments and camera setups entirely different from the training data.
Quantitative Results
On the KITTI Eigen split, the framework achieved state-of-the-art (SOTA) performance with an Absolute Relative (Abs Rel) error of 0.106 for monocular training and 0.102 with monocular + stereo supervision. Furthermore, it demonstrated strong generalization on Make3D with an Abs Rel of 0.306, outperforming existing self-supervised methods.
Qualitative Analysis
One of the most significant advantages of this framework is the visual quality of the predicted depth maps. By effectively harvesting implicit motion cues, the model produces significantly sharper edges around vehicles and pedestrians.
As shown in the comparisons above, previous methods often suffer from "halo" effects or depth bleeding around moving objects. The proposed joint learning approach maintains structural integrity, even in dynamic scenes.
Conclusion and Future Work
By integrating motion features directly into the depth estimation process, this framework overcomes traditional limitations regarding moving objects in self-supervised learning. While the model relies on temporal context (consecutive frames), making it less effective for single-image inference without video, its performance on standard benchmarks sets a new bar for accuracy and contour definition. Future work may involve integrating advanced visual odometry to further refine global pose estimation.
Summary updated