Learning Projection-Aware 360-Degree Image Rectification via Dual-Projection Fusion
This study introduces a dual-stream network to simultaneously estimate camera inclination angles and reconstruct upright 360-degree panoramic images. It achieves this by combining local spatial features from equirectangular projections with global contextual cues from cubemap projections. The method uses adaptive fusion and specialized blocks to ensure accurate geometric alignment, outperforming existing techniques on large datasets like SUN360.