Robust Group Synchronization via Quadratic Programming

Authors: Yunpeng Shi, Cole M Wyeth, Gilad Lerman

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We test our methods for rotation averaging. In 3.1, we describe the implementation details of all tested algorithms. In 3.2 we report the estimation of the corruption levels and rotations on synthetic data generated by UCM for SO(3). In 3.3, we compare the performance of different algorithms on the Photo Tourism dataset (Wilson & Snavely, 2014).
Researcher Affiliation Academia 1Program in Applied and Computational Mathematics, Princeton University 2School of Mathematics, University of Minnesota.
Pseudocode Yes Algorithm 1 DESC-PGD and Algorithm 2 DESC-SO(3) (DESC)
Open Source Code Yes The supplemental code is in https://github.com/ColeWyeth/DESC
Open Datasets Yes For experiments with real data, we used the Photo Tourism dataset, which was introduced in Wilson & Snavely (2014). We compare DESC and other algorithms on synthetic data generated according to UCM, with and without noise.
Dataset Splits No The paper uses synthetic and Photo Tourism datasets but does not provide specific details on training, validation, or test splits (e.g., percentages, sample counts, or explicit statements about standard splits used).
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, processor types, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper mentions using 'default Matlab CVX LP solver' for a baseline comparison and using 'codes provided by the respective papers' for CEMP and MPLS, but it does not provide specific version numbers for its own software dependencies like Python, PyTorch, or other libraries used for DESC.
Experiment Setup Yes For the synthetic data experiments, we ran DESC with a constant step size of 0.01. The maximum number of iterations was set to 100. For real data, due to the large sizes of the datasets, we increased the step size to 1 in order to accelerate the convergence and we decreased the maximum number of iterations to 30. Otherwise, all parameter settings were identical. The number of cycles sampled was chosen as one quarter of the median number of cycles per edge, or at least 30.