Analysis of NaN divergence in training monocular depth estimation model

Bum Jun Kim, Hyeonah Jang, Sang Woo Kim

Abstract: Highlights•We analyze NaN loss during the training of a monocular depth estimation network.•The use of square root loss causes vulnerabilities in optimization.•The log-sigmoid function has numerical stability issues and is prone to NaN loss.•We find that a critical error exists in the implementation of variance computation.•Our guidelines guarantee the absence of NaN loss without performance degradation.

External IDs:doi:10.1016/j.asoc.2025.114051