Reproduce “GenDepth: Generalizing Monocular Depth Estimation for Arbitrary Camera Parameters via Ground Plane Embedding” on OpenResearch

About this paper

Karlo Koledić, Luka Petrović, Ivan Petrović, Ivan Marković

The GenDepth framework addresses the challenge of generalizing monocular depth estimation for autonomous driving by integrating a novel ground plane embedding with adversarial domain alignment. The method demonstrated superior zero-shot cross-dataset metric depth estimation, outperforming state-of-the-art techniques on various real-world benchmarks and significantly enhancing visual odometry performance.

Key insights

• Explicitly embedding camera parameters as a Ground Plane Embedding provides direct, per-pixel geometric cues that significantly improve generalization across varying camera setups.
• A synergistic combination of diverse synthetic data training, adversarial domain alignment for environmental invariance, and GPE for camera parameter awareness is crucial for robust sim2real transfer.
• The model's ability to generalize to unseen continuous camera parameter variations is enhanced by incorporating attention-based feature refinement and explicit geometric encoding.

Summary by alphaXiv.