QML-360°: Permutation-Equivariant Quantum Graph Learning for 360° Image Classification with Distributed Training

Processing 360◦ images for machine vision tasks can significantly enhance various applications in augmented and virtual reality, autonomous driving, and drone surveillance. However, two main challenges arise in this area: the large feature space and angular distortions. To address this, we propose QML-360◦, a hybrid quantum–classical framework for 360◦ image classification that learns on spherical graphs. It uses a permutation-equivariant quantum head to preserve symmetry under node permutations. To address memory and runtime constraints on near-term quantum devices (and in classical simulation), we adopt a distributed training strategy. Gradient evaluations are parallelized across workers, and split-shot estimation reduces per-device load while maintaining estimator quality. Across datasets, our method achieves competitive performance relative to the baselines under large rotation perturbations while reducing per-device compute and memory. In addition, we establish theoretical formulations for equivariance, gradient variance, and cost scaling, and present empirical results that demonstrate their effectiveness.