Cycle-Sync

Cycle-Sync is a robust global camera pose estimation method for recovering camera locations from pairwise directions under heavy corruption. This repository contains parallel Python and MATLAB implementations, synthetic-data generators, real-data demos, and baseline comparisons.

📄 Citation

If you find Cycle-Sync useful in your research, please consider citing our NeurIPS 2025 spotlight paper:

👉 OpenReview PDF

The location solver uses cycle-consistency message passing, T-AAB initialization, and a Welsch-type reweighting scheme. The default parameters are:

Parameter	Default
Maximum iterations	20
Cycle weight beta	20
Annealing schedule	lambda_t = t / (t + 10)
IRLS regularization	1e-8
T-AAB triangle truncation	sinmin = 0.6
Robust weight update	exp(-4h) / (h + delta)

Repository layout

.
├── Cycle-Sync-Python/      # Python implementation, demos, tests, and real-data .mat files
├── Cycle-Sync-Matlab/      # MATLAB implementation, demos, tests, and matching real-data .mat files
└── README.md             # Common entry point

The Python and MATLAB folders are intended to be equivalent at the experiment level: they use the same synthetic corruption models, the same real-data .mat format, the same Cycle-Sync defaults, and the same baseline list.

Included methods

Method	Description
Cycle-Sync	Cycle-consistency reweighted WLS location solver
LUD	Least Unsquared Deviations baseline
ShapeFit	ADMM ShapeFit baseline
BATA	Baseline Desensitizing Translation Averaging
FusedTA	Fused Translation Averaging baseline

Quick start: Python

cd Cycle-Sync-Python
pip install -r requirements.txt
python demo/run_one_synthetic_experiment.py
python demo/run_one_real_dataset.py data/real_precomputed/delivery_area_location.mat --alignment robust
python experiments/compare_methods_real.py --data-dir data/real_precomputed --alignment robust

Run tests:

cd Cycle-Sync-Python
pytest -q tests

Quick start: MATLAB

cd Cycle-Sync-Matlab
startup
run('demo/run_one_synthetic_experiment.m')
run('demo/run_one_real_dataset.m')
run('experiments/compare_methods_real.m')

For real-data experiments, the recommended evaluation uses robust signed-scale/translation alignment. This prevents a few extreme outlier cameras from dominating the fitted scale and translation while still reporting camera errors over the full camera set.

Representative synthetic result

Demo configuration:

model = uniform corruption
n = 100
p = 0.5
q = 0.8
sigma = 0
seed = 2025

Method	Median error	Mean error
Cycle-Sync	1.68e-07	0.158006
LUD	0.950856	0.988628
ShapeFit	1.496418	1.513211
BATA	0.607546	0.756063
FusedTA	0.390215	0.810473

Cycle-Sync reaches near-exact median recovery at 80% uniform corruption in this demo.

Representative real-data result

The small real-data demo set contains two scenes selected because Cycle-Sync has a large separation from the strongest baseline:

data/real_precomputed/delivery_area_location.mat
data/real_precomputed/facade_location.mat

All real-data numbers below use robust signed-scale/translation alignment.

Dataset	Method	Median error	Mean error
delivery_area	Cycle-Sync	0.035242	0.36387
delivery_area	LUD	0.196842	0.338749
delivery_area	ShapeFit	0.261317	0.435902
delivery_area	BATA	0.289026	0.46202
delivery_area	FusedTA	0.823736	0.99406
facade	Cycle-Sync	0.002947	0.234012
facade	LUD	0.05868	0.46065
facade	ShapeFit	0.95011	1.032994
facade	BATA	0.01799	0.545244
facade	FusedTA	0.90159	0.996432

Average median translation error on these two real-data demos:

Method	Avg. median error
Cycle-Sync	0.019095
LUD	0.127761
BATA	0.153508
ShapeFit	0.605714
FusedTA	0.862663

On this two-scene demo, Cycle-Sync has an average median error of 0.019095, compared with 0.127761 for LUD, 0.153508 for BATA, 0.605714 for ShapeFit, and 0.862663 for FusedTA.

Real-data file convention

The real-data loaders expect precomputed .mat files containing:

AdjMat34
tijMat34
t_orig2_cntrd
R_global

Recommended compact demo data:

Cycle-Sync-Python/data/real_precomputed/
├── delivery_area_location.mat
└── facade_location.mat

Cycle-Sync-Matlab/data/real_precomputed/
├── delivery_area_location.mat
└── facade_location.mat

Output files

The demos write result tables and summaries under results/. Typical outputs include:

summary.csv
summary.mat
real_method_by_dataset.csv
real_method_summary.csv
per_camera_errors.csv

License

Cycle-Sync is provided for educational and research use under the license terms included in the LICENSE files.