Unsupervised monocular endoscopic depth estimation and overestimation

This work finds a new phenomenon of depth overestimationd→∞ in unsupervised monocular endoscopic depth prediction. We theoretically analyze the impact of a photometric consistency constraint loss widely used to measure the photometric inconsistency between two continuous endoscopic video frames during training. To resolve this overestimation problem, we propose a new deep unsupervised learning architecture of structure-valid mask transformer with a new unsupervised training strategy to simultaneously predict monocular depth and pose, without annotating any clinical endoscopic data. Specifically, this architecture employs dense convolution and hierarchical transformer to encode multiscale local and global features, and uses residual attention to effectively fuse or decode these features. More interestingly, structure-valid masks are generated by effective motion-driven photo-difference to establish a new hybrid loss function with photometric structure-aware consistency and 3D geometric consistency for unsupervised training of our proposed model. Consequently, our methods can also deal with weak textures (structures) and illumination variations in monocular endoscopic video images. We evaluated our methods on both public synthetic-labeled and in-house clinical endoscopic video images, with the comprehensive experimental results showing that our unsupervised learning methods can attain critically higher depth accuracy than state-of-the-art monocular depth estimation models, as well as more reasonable depth distributions and texture details in predicted depth maps. Our source code has been publicly released on https://github.com/WKFan2026/SVMT.