Is the accuracy–human-likeness trade-off inherent, or can models surpass humans in both accuracy and robust generalization?

Ascertain whether the observed trade-off between metric accuracy and human-likeness in monocular depth estimation is an inherent constraint on robust perception, or whether models can be developed that simultaneously surpass human capabilities in both metric accuracy and robust generalization.

Background

The study finds that the most metrically accurate monocular depth estimators often diverge from human-like error patterns, forming a clear inverse-U relation between accuracy and human similarity. This raises a fundamental issue: whether such divergence is an unavoidable consequence of optimizing for high accuracy in the KITTI domain or a contingent artifact of current methods and datasets.

Resolving whether the trade-off is intrinsic, or instead circumventable with new architectures, training regimes, or data, would shape the strategic direction of human-aligned 3D vision—either toward emulating human perceptual strategies or pursuing alternative routes that achieve both superior accuracy and robust generalization.