Does human-likeness improve out-of-distribution robustness in monocular depth estimation?

Determine whether increasing the similarity of monocular depth estimators to human depth perception—specifically in terms of error pattern alignment—improves out-of-distribution robustness when evaluated on datasets beyond the models’ training distribution.

Background

The paper compares human depth estimates with 69 monocular depth estimation models on KITTI and reports an inverse-U trade-off: as models achieve higher metric accuracy, their error patterns tend to diverge from human judgments. While this work establishes a nuanced relationship between accuracy and human-likeness, it does not test whether human-aligned error structures confer advantages on out-of-distribution data.

The authors highlight that aligning model behavior with human perception is thought to enhance robustness in other domains, but for monocular depth estimation it remains untested whether strengthening human-like characteristics actually leads to better generalization beyond the training distribution. Clarifying this would inform whether human-centric evaluation should be adopted as a robustness proxy in depth estimation.