Self-Supervised Depth and 3D Motion Estimation from Monocular Thermal Video Based on Multispectral Consistency Loss

A study published at ICRA 2022 proposed a self-supervised learning method that simultaneously estimates depth and ego-motion from monocular thermal video. The method uses multispectral consistency consisting of a temperature consistency loss and a photometric consistency loss, and designs a differentiable forward warping module that transforms the thermal camera coordinate system into a visible-light camera, thereby providing a complementary self-supervised signal.