Research map / 3D vision / Depth and stereo
Unsupervised Photometric-Consistent Depth Estimation from Endoscopic Monocular Video
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PC-Depth, an unsupervised monocular depth method for endoscopic video that adds a physics-based Highlight Loss and a Photometric Match module (spotlight illumination model) to handle specular highlights and joint light-camera motion.
Problem
Unsupervised monocular depth estimation in endoscopy fails because photometric consistency across frames is violated by specular highlights (local) and by the simultaneous motion of the light source and camera (global).
Approach
Uses the specular reflection model to add a Highlight Loss that aligns surface normals with the viewing direction in highlight regions, and a Photometric Match module that analytically computes the inter-frame photometric ratio from a spotlight illumination model and applies it to the warped synthetic frame, without optical flow or extra networks.
Key results
- On C3VD, PC-Depth reaches AbsRel 0.076 and RMSE 3.279 vs. 0.081/6.550 for LightDepth, the strongest prior in that table (best in block).
- On SCARED, PC-Depth attains AbsRel 0.057 and RMSE 5.362, improving over Monodepth2* (0.063/5.695) and DepthAnything (0.068/5.643).
- On SERV-CT (trained on SCARED, zero-shot), PC-Depth achieves AbsRel 0.104 vs. 0.116 for SC-Depth, the best prior.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PC-Depth instead of LightDepth | PC-Depth · C3VD | AbsRel: 0.081 -> 0.076 | Helped |
| use PC-Depth instead of Yang et al. | PC-Depth · SCARED | AbsRel: 0.062 -> 0.057 | Helped |
| use PC-Depth instead of SC-Depth | PC-Depth · SERV-CT | AbsRel: 0.116 -> 0.104 | Helped |
| add Smoothness Loss with Highlight Removal (LS) (ablation) | PC-Depth with LH + LP · C3VD | AbsRel: 0.082 -> 0.076 | Helped |
| add Photometric Loss with Photometric Match (LP) (ablation) | PC-Depth with LH only · C3VD | AbsRel: 0.091 -> 0.082 | Helped |
| add Highlight Loss (LH) (ablation) | PC-Depth with LP only · C3VD | AbsRel: 0.079 -> 0.082 | Hurt |
| add Smoothness Loss with Highlight Removal (LS) (ablation) | PC-Depth with LH only · C3VD | AbsRel: 0.091 -> 0.091 | No clear effect |
Surprising findings
- Ego-motion improvement is less pronounced than depth improvement, which the authors attribute to pixel-level mapping being more sensitive to photometric inconsistency.
Limitations
- Photometric Match relies on a Lambertian surface assumption, and highlight regions are excluded from the photometric loss.
- Generalization is only tested on SERV-CT with models trained on SCARED.
- Ego-motion gains are modest compared with depth gains.
Experimental setup
Datasets: C3VD SCARED
Benchmarks: C3VD SCARED SERV-CT
Metrics: AbsRel SqRel RMSE RMSE log threshold accuracy (δ<1.25, δ<1.25^2) ATE trans ATE rot
Baselines: Monodepth2 SC-Depth Endo-SfM AF-SfM LightDepth MonoLoT DepthAnything SfMLearner MVDE Yang et al. 2024b
Builds on: SC-Depth Monodepth2 Endo-SfM Arnold et al. highlight detection Modrzejewski spotlight illumination model
Closest papers by meaning
- Manifold-Guided Stereo-Monocular Refinement for Endoscopic Stereo Disparity Estimation (NeurIPS 2026 · similarity 0.88)
- Depth Any Endoscopy: Towards Self-Supervised Generalizable Depth Estimation in Monocular Endoscopy (CVPR 2026 · similarity 0.88)
- DeLightMono: Enhancing Self-Supervised Monocular Depth Estimation in Endoscopy by Decoupling Uneven Illumination (AAAI 2026 · similarity 0.88)
- PolypSense3D: A Multi-Source Benchmark Dataset for Depth-Aware Polyp Size Measurement in Endoscopy (NeurIPS 2025 · similarity 0.84)
- Video Depth Anything: Consistent Depth Estimation for Super-Long Videos (CVPR 2025 · similarity 0.84)
- Depth Pro: Sharp Monocular Metric Depth in Less Than a Second (ICLR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.