MADPose

This repo contains the official implementation of the solvers and estimators proposed in the paper "Relative Pose Estimation through Affine Corrections of Monocular Depth Priors" (CVPR 2025 Highlight). The solvers and estimators are implemented using C++, and we provide easy-to-use Python bindings.

Note: "MAD" is an acronym for "Monocular Affine Depth".

Overview: Our method takes a pair of images as input, runs off-the-shelf feature matching and monocular depth estimation, then jointly estimates the relative pose, scale and shift parameters of the two depth maps, and optionally the focal lengths.

📢 News

• 2025-02-27: Our paper is accepted to CVPR 2025! See you in Nashville, TN!

Evaluation

We develop three solvers for relative pose estimation that explicitly account for independent affine (scale and shift) ambiguities, tailored for three setups of cameras: calibrated, shared-focal, and unknown focal lengths (two-focal). The solvers are further combined with classic point-based solvers and epipolar constraints in our hybrid RANSAC estimators. Our estimators show consistent improvements over classic point-based methods (e.g. PoseLib) and recent two-view reconstruction methods DUSt3R and MASt3R, across different datasets with different feature matchers and monocular depth estimation models.

Some highlight results (probably best viewed in light theme on GitHub!):

Pose Error AUCs on ScanNet-1500 with shared-focal setting

Matches	Method	MD Model	AUC@5°	AUC@10°	AUC@20°
SP+SG	PoseLib-6pt	-	12.84	28.13	45.64
Ours-sf	DA-met.	18.35	37.54	57.58
RoMa	PoseLib-6pt	-	27.17	49.24	67.42
Ours-sf	DA-met.	29.81	53.11	71.15
MASt3R	PoseLib-6pt	-	30.28	54.16	72.87
Ours-sf	DA-met.	31.87	56.20	74.51
MASt3R	32.58	56.99	74.91
Reference entry - DUSt3R	25.90	48.45	68.03
Reference entry - MASt3R	23.94	46.44	66.18

Pose Error AUCs on 2D-3D-S sampled image pairs with unknown focal lengths

Matches	Method	MD Model	AUC@2°	AUC@5°	AUC@10°	AUC@20°
SP+LG	PoseLib-7pt	-	5.85	13.95	21.94	30.71
Ours-tf	DAv2-met.	9.15	22.22	32.80	43.26
RoMa	PoseLib-7pt	-	8.73	20.31	30.45	41.48
Ours-tf	DAv2-met.	13.50	29.19	42.18	54.42
MASt3R	PoseLib-7pt	-	12.58	30.27	45.57	59.85
Ours-tf	DAv2-met.	18.05	39.92	56.64	70.86
MASt3R	22.44	48.02	64.79	76.55
Reference entry - DUSt3R	6.43	24.47	42.39	58.36
Reference entry - MASt3R	13.39	38.41	57.92	71.91

Please refer to the paper for more results and discussions.

Installation

Install from PyPI

We are working on setting up wheel for easy installation using PyPI. Currently please use the following method to install from source.

Install from source

Install dependencies

sudo apt-get install libeigen3-dev libceres-dev libopencv-dev

Note: The two-focal estimator currently relies on cv::recoverPose from OpenCV, we plan to remove dependency on OpenCV in future updates.

Clone the repo

git clone --recursive https://github.com/MarkYu98/madpose

Build and install the Python bindings

pip install .

If you would like to see the building process (e.g. CMake logs) you can add -v option to the above command.

Note: MADPose also depends on PoseLib. By default, CMake will automatically build PoseLib using FetchContent. Set FETCH_POSELIB CMake option to OFF if you prefer to use a self-installed version.

Check the installation

python -c "import madpose"

You should not see any errors if MADPose is successfully installed.

Usage

Estimators

We provide Python bindings of our 3 hybrid estimators for image pairs with calibrated cameras, shared-focal cameras, and cameras with unknown focal lengths (two-focal).

The estimators take HybridLORansacOptions and EstimatorConfig for related settings, some useful settings are: ```python import madpose

options = madpose.HybridLORansacOptions() options.min_num_iterations = 100 options.max_num_iterations = 1000 options.success_probability = 0.9999 options.random_seed = 0 # for reproducibility options.final_least_squares = True options.threshold_multiplier = 5.0 options.num_lo_steps = 4

squared px thresholds for reprojection error and epipolar error

options.squared_inlier_thresholds = [reproj_pix_thres ** 2, epipolar_pix_thres ** 2]

weight when scoring for the two types of errors

options.data_type_weights = [1.0, epipolar_weight]

est_config = madpose.EstimatorConfig()

if enabled, the input min_depth values are guaranteed to be positive with the estimated depth offsets (shifts), default: True

est_config.min_depth_constraint = True

if disabled, will model the depth with only scale (onl