Reproduce “MonoProb: Self-Supervised Monocular Depth Estimation with Interpretable Uncertainty” on OpenResearch

About this paper

Rémi Marsal, Florian Chabot, Angelique Loesch, William Grolleau, Hichem Sahbi

MonoProb presents a self-supervised method for monocular depth estimation that provides both depth predictions and interpretable uncertainty directly corresponding to expected depth error. The framework achieves competitive depth accuracy on KITTI while substantially improving uncertainty quality, as measured by newly introduced absolute evaluation metrics like Absolute Relative Uncertainty (ARU) and Root Mean Square Uncertainty Error (RMSU).

Key insights

• Interpretable uncertainty, directly correlating with expected depth error, can be achieved in self-supervised monocular depth estimation within a single forward pass.
• Reformulating the photometric reconstruction loss within a probabilistic framework allows the network to learn both depth and its associated uncertainty end-to-end.
• New absolute uncertainty metrics (ARU, RMSU) are essential for evaluating whether predicted uncertainty magnitudes truly correspond to actual depth errors, providing a more robust assessment than relative ranking metrics.

Summary by alphaXiv.