PanoLOG: The First Large-Scale Outdoor 3D Reconstruction System for Panoramic Input

Weijian Chen (formerly an algorithm intern at Insta360 Research Institute) and Lu Qi present PanoLOG, a novel large‑scale outdoor 3D reconstruction pipeline designed for panoramic inputs. The work is accompanied by the open‑source Pano360 benchmark.

Traditional city‑scale reconstruction relies on massive narrow‑FoV photos and block‑wise training, which incurs high acquisition and computation costs. While panoramic cameras can capture a full 360° view in a single shot, their omnipresent visibility invalidates the conventional spatial‑visibility partitioning strategy.

PanoLOG addresses this contradiction by shifting the partitioning criterion from spatial visibility to gradient contribution, thereby restoring block‑parallel training under panoramic conditions and achieving superior rendering quality with a smaller model footprint.

Two‑Stage Framework

Stage 1 performs a global coarse optimization on all panoramas to obtain a stable geometric prior. It introduces explicit sky gaussians (100 k sky gaussians placed on a sphere ten times the scene scale, with frozen position gradients) and panoramic depth supervision using the ERP‑compatible DAP depth model, whose loss weight decays exponentially to transition from geometry‑driven to photometric‑driven optimization.

Stage 2 refines each block with the G²PS (GGPS) method, which combines (1) geometry‑based spatial partitioning—adaptive AABB construction based on camera distribution and triangulation reliability, followed by balanced uniform voxelization; and (2) gradient‑based camera‑block assignment—after Stage 1 convergence, average gradient magnitudes per camera‑block are computed, and a camera is assigned to a block if it lies inside the block or its normalized gradient score exceeds 0.8. Block optimization then re‑initializes gaussians from Stage 1, freezes sky gaussians, and uses periodic opacity resets to prune inactive gaussians while preserving cross‑block consistency.

The final rendering adopts ERP projection, mapping 3D points to pixel coordinates and projecting 3D covariances to 2D via the ERP Jacobian, which correctly handles the 360° field of view.

Experiments

All methods are trained for 30 000 steps with two partitions on a single NVIDIA RTX 4090 (24 GB). Baselines include large‑scale 3DGS methods H3DGS, CityGaussian, DOGS, Momentum‑GS (ERP→cubemap), as well as panoramic‑specific methods OmniGS, ODGS, SpaGS and a cubemap 3DGS baseline. Metrics are PSNR, SSIM, LPIPS and model size.

On the self‑built Pano360 A1 drone scenes (NSC, NSK), PanoLOG achieves the best PSNR/SSIM/LPIPS on NSC and the best SSIM/LPIPS on NSK, while using the smallest model. On the Pano360 X5 handheld street scenes (BAX, NSN), it improves PSNR by 0.64 dB (BAX) and 1.16 dB (NSN) and reduces model size to 1/2.9–1/7.5 of the strongest baseline, whose model volume grows to 5.7–7.6 GB. On public datasets Ricoh360 and 360Roam, PanoLOG attains the top scores on all six evaluation items, confirming its generalization.

Outlook

By converting the partitioning assumption from visibility to gradient contribution, PanoLOG enables high‑quality, low‑cost city‑scale reconstruction from panoramic captures and aligns with Insta360’s vision of autonomous imaging agents (Cameraman) that perceive and interact with real environments.