GEDepth: Ground Embedding for Monocular Depth Estimation

Abstract

Monocular depth estimation (MDE) is a fundamental yet ill-posed challenge in computer vision, as a single 2D image can correspond to infinite 3D configurations. Current state-of-the-art models often act as "black boxes," overfitting to specific camera geometries (height, tilt, focal length) and failing when deployed on different platforms. We introduce GEDepth, a lightweight, plug-and-play module that explicitly embeds camera geometry into the depth learning process. By decoupling camera parameters from visual cues and introducing an adaptive ground embedding, GEDepth significantly improves accuracy and cross-domain generalization.

Introduction: The Geometry Gap in MDE

The primary hurdle in monocular depth estimation is the inherent ambiguity of scale. Models trained on a specific vehicle-mounted camera often rely on learned priors—such as the expected height of a sedan—to estimate distance. However, if the camera's height or tilt changes, these visual cues become misleading. GEDepth addresses this by providing the model with a geometric "floor": a theoretical ground plane calculated from known camera intrinsics and extrinsics.

Methodology

GEDepth is designed to be integrated into existing depth estimation backbones (e.g., BTS, DepthFormer). The module operates through three synergistic components:

1. Ground Depth Formulation: The module uses camera intrinsics and extrinsics to mathematically derive the theoretical depth of every pixel, assuming a perfectly flat ground plane.
2. Adaptive Ground Embedding: To account for real-world terrain like hills and bumps, the "Adaptive" mode predicts a Ground Slope Map. This allows the geometric reference to dynamically adjust to undulations.
3. Ground Attention: A lightweight attention mechanism learns to distinguish between ground pixels and object pixels. It applies the geometric ground prior to the road while relying on standard visual feature extraction for objects like vehicles or pedestrians.

Expand

Source, p. 3

This architecture allows the model to learn ground identification implicitly through standard depth loss, requiring no additional semantic labels.

Key Insights and Results

A significant finding of this work is the emergence of Implicit Segmentation. The Ground Attention Map naturally learns to mask out non-ground objects without explicit supervision. This allows the network to anchor its scale on the geometrically stable ground while focusing its representational power on complex objects.

Expand

Source, p. 5

Performance and Generalization

GEDepth consistently enhances the performance of various backbone architectures:

• KITTI Benchmark: Integrating GEDepth with DepthFormer reduced the Absolute Relative Error (Abs Rel) from 0.052 to 0.048, setting a new state-of-the-art.
• Cross-Domain Robustness: In zero-shot testing (training on KITTI and testing on the more diverse DDAD dataset), GEDepth-enhanced models outperformed baselines by a massive margin, reducing Abs Rel from 0.644 to 0.261.
• Resolution Stability: Unlike standard models whose accuracy fluctuates with input resolution, GEDepth remains stable by explicitly accounting for changes in effective focal length.

Conclusion and Future Work

By providing a solid geometric foundation, GEDepth transforms monocular depth estimation from a pure "guessing" game into a grounded geometric inference task. While the model currently excels on typical road grades, future iterations may focus on more extreme terrain or extending the "ground" concept to indoor environments (floors and walls). GEDepth proves that even in the era of deep learning, explicit geometric priors remain a powerful tool for building robust and generalizable vision systems.

@misc{ma2023gedepth,
  author = {Zhuang Ma},
  title = {{GEDepth: Ground Embedding for Monocular Depth Estimation}},
  year = {2023},
  doi = {10.1109/iccv51070.2023.01168},
  howpublished = {OpenReview},
}