Intrinsic geometry
Model features in relation to the spherical domain of panoramic imagery.
SO3UFormer.
Learning Intrinsic Spherical Features for Rotation-Robust Panoramic Dense Prediction — Qinfeng Zhu, Yunxi Jiang and Lei Fan.
Project premise
Panoramic images live on a sphere. Their representations should respect that geometry.
SO3UFormer studies intrinsic spherical features for panoramic dense prediction, with an explicit focus on robustness to rotation. The project connects the geometry of the observation domain to the learned representation rather than treating a panorama as an ordinary planar image.
Rotation robustness
Design for stable dense prediction when the spherical view rotates.
Dense understanding
Connect sphere-native representation learning to pixel-level panoramic tasks.