Uncertainty-aware probabilistic streaming depth estimation for endoscopic video

Monocular depth estimation from endoscopic video supports 3D reconstruction, localization, and surface measurement in minimally invasive procedures. Existing streaming methods improve temporal consistency but remain deterministic, lacking explicit uncertainty awareness for ambiguous regions caused by reflection, weak texture, or blur. This work introduces an uncertainty-aware probabilistic framework for streaming endoscopic depth estimation that jointly predicts dense depth maps and per-pixel aleatoric uncertainty. The model uses a temporal depth backbone with Mamba modules and a parallel uncertainty branch, optimized via an annealed heteroscedastic objective. Experiments on C3VD and SimCol3D benchmarks show state-of-the-art performance on multiple metrics. The code is publicly available at https://github.com/HaoLi12345/uncertaintymde.