Monocular depth and pose estimation play an important role in the development of colonoscopy-assisted navigation, as they enable improved screening by reducing blind spots, minimizing the risk of missed or recurrent lesions, and lowering the likelihood of incomplete examinations. However, this task remains challenging due to the presence of texture-less surfaces, complex illumination patterns, deformation, and a lack of in-vivo datasets with reliable ground truth. In this paper, we propose PRISM (Pose-Refinement with Intrinsic Shading and edge Maps), a self-supervised learning framework that leverages anatomical and illumination priors to guide geometric learning. Our approach uniquely incorporates edge detection and luminance decoupling for structural guidance. Specifically, edge maps are derived using a learning-based edge detector (e.g., DexiNed or HED) trained to capture thin and high-frequency boundaries, while luminance decoupling is obtained through an intrinsic decomposition module that separates shading and reflectance, enabling the model to exploit shading cues for depth estimation. Experimental results on multiple real and synthetic datasets demonstrate state-of-the-art performance. We further conduct a thorough ablation study on training data selection to establish best practices for pose and depth estimation in colonoscopy. This analysis yields two practical insights: (1) self-supervised training on real-world data outperforms supervised training on realistic phantom data, underscoring the superiority of domain realism over ground truth availability; and (2) video frame rate is an extremely important factor for model performance, where dataset-specific video frame sampling is necessary for generating high quality training data.
Figures & tables
Figure 1: An overview of our proposed self-supervised monocular depth estimation framework. Our key contribution is to explicitly guide the DepthNet with priors from two parallel branches: a LumNet, inspired by IID-SfMLearner, to extract an illumination-aware luminance map ( Lt ), and an EdgeNet, based on DexiNed, to extract a high-fidelity structural edge map ( Et ). The inputs to the DepthNet are the concatenation of the raw image It , the luminance map Lt , and the edge map Et .
Model
RMSE ↓
LogRMSE ↓
MAE ↓
MedAE ↓
AbsRel ↓
SqRel ↓
ATE ↓
Monodepth2 [ 11 ]
6.3138
0.1841
4.3699
3.1279
0.1501
1.0168
2.7025
MonoViT [ 12 ]
6.8544
0.1789
4.6146
3.0384
0.1473
1.0958
2.4298
AF-SfM _pretrained [ 19 ]
17.3395
0.5216
12.0918
8.2881
0.4126
7.2491
-
AF-SfM _finetuned [ 19 ]
9.9636
0.2658
6.3069
3.4845
0.2165
3.7232
4.4841
IID-SfM _pretrained [ 8 ]
17.6461
0.5269
12.3394
8.4590
0.4178
7.4915
-
IID-SfM _finetuned [ 8 ]
8.6977
0.2552
5.6891
3.5772
0.1996
1.9452
3.1358
Table 1: Quantitative comparison of depth estimation on the C3VD test set for models trained with Hyper-Kvasir. Best results are in bold , second best are underlined .
Figure 2: Qualitative evaluation of generalization on real-world endoscopic video frames with challenging improvements boxed in red.
Figure 3: Qualitative comparison of estimated camera trajectories across four models on three representative sequences from the C3VD dataset. Each 3D plot shows the predicted path color-coded by temporal order.
Models
Pose Edge
Pose Lum
Depth Edge
Depth Lum
RMSE ↓
Log RMSE ↓
MAE ↓
MedAE ↓
Abs Rel ↓
Sq Rel ↓
δ<1.25 ↑
δ<1.252 ↑
δ<1.253 ↑
Baseline
–
–
–
–
6.2258
0.1845
4.3176
3.0993
0.1523
1.0360
0.7727
0.9741
0.9961
PE
✓
–
–
–
6.2501
0.1824
4.3260
2.9790
0.1505
1.0610
0.7852
0.9665
0.9945
PL
–
✓
–
–
6.7658
0.2072
4.8068
3.5452
0.1719
1.2877
0.7318
0.9489
0.9910
DE
–
–
✓
–
6.6625
0.2111
4.7625
3.4991
0.1755
1.3095
0.7338
0.9415
0.9891
DL
–
–
–
✓
6.7540
0.2005
4.7099
3.4705
0.1649
1.2652
0.7468
0.9576
0.9951
DEPL
–
✓
✓
–
6.9001
0.2105
4.9384
3.5381
0.1739
1.3404
0.7274
0.9483
0.9896
Table 2: Quantitative comparison of depth estimation with models trained on Hyper-Kvasir and tested on the C3VD test set. Best results are in bold and second best results are in underlined .
Figure 4: Qualitative comparison of ablation variants on four subsequences from the EndoMapper dataset. Each row shows a sample frame and the corresponding predicted depth maps from different prior configurations.
Model
Edge Loss
Pose Estimation
Depth Estimation
ATE ↓
RTE ↓
ROT
RMSE ↓
Log RMSE ↓
MAE ↓
MedAE ↓
AbsRel ↓
SqRel ↓ ↓
Monodepth2 [ 11 ]
–
2.7025
0.2004
0.1250
6.3138
0.1841
4.3699
3.1279
0.1501
1.0168
✓
1.6293
0.2121
0.1286
6.8043
0.1991
4.9267
3.7038
0.1681
1.2604
DLPE
–
2.5101
0.2005
0.1408
6.1246
0.1830
4.3300
3.2582
0.1475
1.0010
✓
2.5858
0.2015
0.1412
6.5946
0.1761
4.4031
2.9848
0.1388
1.0138
✓ †
1.9077
0.2087
0.1496
6.1246
0.1830
4.3300
3.2582
0.1475
1.0010
Table 3: Depth and pose estimation results of models trained with Hyper-Kvasir and tested on C3VD test set with different combinations of structural priors and edge-guided loss. Best configuration per model group is in bold .
Figure 5: Qualitative trajectory comparison under modality-only vs. modality + edge-guided SSIM training on cecum_t4b , sigmoid_t3b and trans_t4b .
Model
Training Dataset
Interval
RMSE ↓
RMSE log ↓
MAE ↓
MedAE ↓
Abs Rel ↓
Sq Rel ↓
δ<1.25↑
δ<1.252↑
δ<1.253↑
Monodepth2 [ 11 ]
C3VD
1
15.1642
0.4196
10.1482
6.3000
0.3047
4.7856
0.4516
0.7353
0.8783
10
7.5960
0.2215
5.0917
3.4709
0.1732
1.4608
0.7161
0.9253
0.9827
20
7.7644
0.2473
5.4523
3.9334
0.2002
1.7340
0.6668
0.8976
0.9725
30
7.7282
0.2403
5.2392
3.6111
0.1878
1.6480
0.6919
0.8956
0.9696
Hyper-Kvasir
1
6.3138
0.1841
4.3699
3.1279
0.1501
1.0168
0.7754
0.9720
0.9956
10
40.5039
1.3728
32.2524
26.0333
1.5865
93.1729
0.1331
0.2595
0.3706
Table 4: Quantitative depth estimation results across models, datasets, and frame intervals. All models are trained on either C3VD or Hyper-Kvasir and always tested on C3VD test data.
Figure 6: Qualitative generalization performance of models trained with Hyper-Kvasir on EndoMapper real dataset. This evaluation highlights model robustness to real-world challenges like specular reflections and complex tissue folds, where Hyper-Kvasir-trained models consistently show an advantage.
Model
RMSE ↓
LogRMSE ↓
MAE ↓
MedAE ↓
AbsRel ↓
SqRel ↓
δ<1.25 ↑
δ<1.252 ↑
δ<1.253 ↑
Self-supervised
7.5960
0.2215
5.0917
3.4709
0.1732
1.4608
0.7161
0.9253
0.9827
Supervised (+GT)
8.2805
0.2443
5.6020
3.7030
0.1921
1.7754
0.6797
0.9022
0.9721
Table 5: Comparison between self-supervised and supervised Monodepth2 [ 11 ] models on C3VD (interval = 10). Supervised model adds scale-consistent depth loss with a weight of 0.5. The better one is marked in Bold
Model
Training Set
Interval
RMSE ↓
LogRMSE ↓
MAE ↓
MedAE ↓
AbsRel ↓
SqRel ↓
PRISM
C3VD
1
14.4647
0.4165
9.7689
6.3846
0.3221
4.7946
C3VD
30
7.7530
0.2460
5.4256
3.9831
0.1972
1.6139
C3VD_supervised
30
8.2836
0.2523
5.6774
3.8749
0.1976
1.7205
Hyper-phasir
1
6.1246
0.1830
4.3300
3.2582
0.1475
1.0010
Table 6: Comparison of depth estimation across datasets, temporal intervals, and supervision modes using PRISM with DLPE settings.
Depth estimation is a foundational component for 3D reconstruction in minimally invasive endoscopic surgeries. However, existing monocular depth estimation techniques often exhibit limited performance to the varying illumination and complex textures of the surgical environment. While applying foundation models offers a promising approach to enhance the depth estimation performance, the domain gap between the natural images used for pre-training and the target endoscopic images leads to significant semantic perception deficiencies. In this study, EndoUFM is introduced as an unsupervised monocular depth estimation framework that innovatively Utilizes dual Foundation Models for Endoscopic images, thereby enhancing the depth estimation performance by leveraging the powerful pre-learned priors. The framework features a novel adaptive fine-tuning strategy that incorporates Random Vector Low-Rank Adaptation (RVLoRA) to enhance model adaptability, and a Residual block based on Depthwise Separable Convolution (Res-DSC) to improve the capture of fine-grained local features. A mask-guided smoothness loss is also introduced to enforce depth consistency within anatomical structures. Extensive experiments on the SCARED, Hamlyn, SERV-CT, and EndoNeRF datasets confirm that our method achieves state-of-the-art performance while maintaining an efficient model size. This work contributes to augmenting surgeons' spatial perception during minimally invasive procedures, thereby enhancing surgical precision and safety, with crucial implications for augmented reality and navigation systems. Our code is available at https://github.com/RealMindyY/EndoUFM.
Xinning Yao, Bo Liu, Bojian Li +3
Image Processing Center, Beihang University, Beijing, 100191, China. · State Key Laboratory of High-Efficiency Reusable Aerospace Transportation Technology, Beijing 102206, China.
Monocular depth estimation (MDE) for colonoscopy is hampered by the domain gap between simulated and real-world images. Existing image-to-image translation methods, which use depth as a posterior constraint, often produce structural distortions and specular highlights by failing to balance realism with structure consistency. To address this, we propose a Structure-to-Image paradigm that transforms the depth map from a passive constraint into an active generative foundation. We are the first to introduce phase congruency to colonoscopic domain adaptation and design a cross-level structure constraint to co-optimize geometric structures and fine-grained details like vascular textures. In zero-shot evaluations conducted on a publicly available phantom dataset, the MDE model that was fine-tuned on our generated data achieved a maximum reduction of 44.18% in RMSE compared to competing methods. Our code is available at https://github.com/YyangJJuan/PC-S2I.git.
Juan Yang, Yuyan Zhang, Han Jia +2
College of Computer Science, Sichuan University, Chengdu, China · Department of Gastroenterology and Hepatology, West China Hospital, Sichuan University, Chengdu, China
Geometric estimation including depth estimation and scene reconstruction is a crucial technique for colonoscopy which can provide surgeons with 3D spatial perception and navigation. However, geometric ground truth in colonoscopy is difficult to obtain due to narrow and enclosed space of the colon, while there is a large feature gap between simulated data and realistic data caused by artifacts and illumination. In this paper, we present CoGE, a novel framework for online monocular geometric estimation during colonoscopy. Firstly, we propose an illumination-aware supervision module based on the Retinex theory to address illumination diversity in different colonoscopy scenes. Moreover, a structure-aware perception module is proposed based on wavelet decomposition to extract common structural and local features of the colon. Both quantitative and qualitative results demonstrate that the proposed model solely trained on simulated data achieves state-of-the-art performance in geometric estimation for both simulated and realistic scenes.
Liangjing Shao, Beilei Cui, Hongliang Ren
Department of Electronic Engineering, The Chinese University of Hong Kong, Hong Kong SAR, China · Shenzhen Loop Area Institute, China