MetricCol: Absolute-Scale Depth and Pose Estimation in Colonoscopy via Geometric Consistency and Domain Adaptation

Abstract

Accurate metric depth and pose estimation are critical for colonoscopic navi-gation and lesion localization. However, existing methods often struggle with scale ambiguity and domain gaps between synthetic and real datasets. To ad-dress these issues, we propose a novel framework consisting of two stages: 1) a fully supervised depth estimation model utilizing synthetic data with anatomi-cal priors to bridge the domain gap between synthetic and real datasets, and 2) a weakly supervised joint learning approach combining camera-aware depth scaling with uncertainty-driven pseudo-labeling to refine metric depth and pose estimation. We validate the framework on both synthetic and real colon-oscopy datasets, achieving superior performance in metric depth (RMSE=3.5408) and pose estimation (average ATE=0.6143). Experimental re-sults on both synthetic and real colonoscopy datasets show superior perfor-mance, robustness under challenging conditions, and demonstrate the clinical applicability of our method.

Results

Table 2. Quantitative depth comparison on C3VD dataset.

Method	Abs Rel(↓)	Sq Rel(↓)	RMSE(↓)	RMSE log(↓)	a1(↑)
DenseDepth	0.1803	2.0154	7.1911	0.2803	0.7183
BTS	0.1534	1.2341	7.1054	0.1954	0.7441
DepthAnythin	0.1136	0.5841	4.5747	0.1368	0.9023
Ours (Stage1)	0.1105	0.5632	4.2915	0.1347	0.9040
Endo-SfMLearner	0.5490	8.2250	12.7652	0.5186	0.3469
AF-SfMLearner	0.2813	3.7943	12.1886	0.3549	0.4674
EndoDAC	0.3432	6.1716	14.3523	0.4136	0.4159
Ours(Baidu Netdisk)	0.0953	0.4019	3.5408	0.1158	0.9429

提取码：jw3M

Figure 1. Depth results comparison on the C3VD dataset.

Table 3. Quantitative pose comparison on C3VD dataset.

Method	t3v1 APE (↓)	t3v1 ATE (↓)	t4v1 APE (↓)	t4v1 ATE (↓)	s3v2 APE (↓)	s3v2 ATE (↓)
ES	0.7014	0.4902	10.740	9.6029	5.2610	4.8089
AF	0.5972	0.4174	5.0222	4.2030	5.1370	4.4748
ED	0.2586	0.2227	1.7911	1.6432	1.7673	1.5261
Ours	0.1486	0.1317	1.2176	1.1370	0.6735	0.5741

Figure 2. Pose results comparison on the C3VD dataset.

Initialization

Create an environment with conda:

conda env create -f conda.yaml
conda activate depth-anything

Install required dependencies with pip:

pip install -r requirements.txt

Download pretrained model from: depth_anything_vitb14. Create a folder named pretrained_model in this repo and place the downloaded model in it.

Dataset

C3VD Dataset.

Download dataset from :c3vd.For the training set, we selected the following samples: c2v2, c3v1, c1v1, t1v2, t3v2, c1v2, s3v1, t4v2, t1v1, t2v1, c2v3, t2v2, and s2v1. The validation set in-cludes: c4v2, c2v1, t2v3, and s1v1. The test set contains: t3v1, c4v1, and d4v1.

SimCol Dataset.

Download dataset from:simcol. The training set consists of 25,942 images, the validation set includes 2,882 imag-es, and the test set contains 9,009 images

Usage

Training

The code will be released as soon as possible.

Evaluation

python evaluate_depth.py --data_path <your_data_path> --eval_split <dataset> --load_weights_folder <the path of loaded weights> --visualize_depth
python evaluate_c3vd_pose.py --data_path <your_data_path> --load_weights_folder <the path of loaded weights> --eval_split <dataset>

Acknowledgment

Our code is based on the implementation of AF-SfMLearner, Depth-Anything, EndoDAC. We thank their excellent works.