Panoramas defeat block-wise 3DGS; PanoLOG repartitions large scenes by render-loss gradient

Geometry and Gradient-based Partitioning for Panoramic Outdoor Reconstruction

Weijian Chen, Weibo Yao, Yuhang Zhang, Xiaolin Tang, Guo Wang, Weijun Zhang, Xitong Gao, Yihao Chen, Hongde Qin, Lu Qi

cs.CV

2026-07-10

PanoLOG repartitions panoramic 3DGS by render-loss gradient instead of frustums, fixing block-wise training collapse. Ships Pano360, the first large-scale panoramic benchmark.

What problem this solves

3D Gaussian Splatting (3DGS) represents a scene as a set of 3D Gaussians with opacity and renders by projecting and compositing them. It trains fast and renders fast, so it has become the default route for scene reconstruction. But scaling it to large outdoor scenes (block-, campus-sized) hurts on both data and compute. A pinhole camera frame covers a narrow field of view, so covering a large area means capturing a huge number of frames.

Panoramic images (using equirectangular projection, ERP, to wrap a full 360° view into one rectangle) look like the fix: one frame does the job of several. But drop-in block-wise methods such as VastGaussian and CityGaussian stop working. Their partitioning logic is built on the pinhole camera frustum, the cone a camera can see, and they cut a scene into blocks by which cameras predominantly see which volume. A panorama sees everywhere at once, so every spatial region is coupled across every viewpoint; the frustum-based cut loses all discriminative power, and block-wise optimization degenerates into global training. The whole point of partitioning, parallelism and memory savings, disappears.

PanoLOG, from Insta360 Research with several universities, targets exactly that failure and redesigns partitioning so it works on panoramas.

Method

PanoLOG is a two-stage coarse-to-fine framework.

Stage I is global coarse training: optimize jointly over all panoramas to build stable geometric priors. Unlike prior work that supervises only with photometric loss (aligning rendered pixels to the input image), it adds panoramic monocular depth supervision from the start and models a dedicated sky sphere.

Sky-sphere modeling. Sky regions carry no valid SfM (structure-from-motion) geometry, so the near-field Gaussians drift and produce floater artifacts. PanoLOG places a set of dedicated sky Gaussians at 10× the scene radius, initialized by uniform spherical sampling, and freezes them during block-wise refinement to keep cross-block views consistent.

Panoramic monocular depth supervision. ERP projection stretches pixels at the poles, which weakens SfM triangulation. PanoLOG uses Depth-Any-Panorama to produce monocular inverse depth maps, aligned to sparse SfM depth via per-image affine parameters. The depth loss carries an exponentially decaying weight (0.5 down to 0.01): strong geometric guidance early, then released, so depth-estimation error does not cap final quality.

Stage II is block-wise refinement, using G²PS (Geometry and Gradient-based Partitioning Strategy) for spatial partitioning and camera allocation.

Geometry-based partitioning. Panoramic scenes are unbounded, so a uniform grid gives unbalanced load. G²PS expands the camera-trajectory bounds based on triangulation reliability to build an adaptive axis-aligned bounding box (AABB): margin = median nearest-neighbor camera distance × triangulation range factor. This turns an unbounded outdoor environment into a stable reconstruction region.

Gradient-based camera allocation. This is the step that addresses frustum failure. Since geometry cannot use a frustum to decide which block a camera belongs to, G²PS scores each camera's contribution to each block from the rendering-loss gradient. If gradient importance exceeds a threshold (τgrad=0.8), the camera is assigned to the block even if it physically sits in a neighboring one, so viewpoints with significant observational contribution are not dropped.

Results

On the new Pano360 benchmark, PanoLOG is compared against four block baselines: H3DGS, CityGaussian, DOGS, Momentum-GS.

Drone scenes (NSC, NSK):

MethodNSC PSNRNSK PSNRNSK SSIMNSK LPIPS
H3DGS27.7824.150.81540.1934
CityGaussian27.7324.830.81760.1983
PanoLOG28.1824.640.82430.1916

Handheld scenes (BAX, NSN) show bigger gains: +0.64 dB and +1.16 dB over the strongest baseline, with model size only 1/2.9 to 1/7.5 of H3DGS.

MethodBAX PSNRNSN PSNR
H3DGS20.7123.45
CityGaussian19.8421.95
PanoLOG21.3524.61

On public panoramic datasets Ricoh360 and 360Roam, against OmniGS, ODGS, SpaGS and cubemap 3DGS, PanoLOG is best on every metric (Ricoh360 PSNR 26.48, 360Roam 25.83).

The ablation (NSK) shows all three components pull weight: full model 24.65 PSNR, dropping G²PS falls to 24.14, dropping the sky sphere to 24.15, dropping depth supervision to 24.61.

Why it matters

Large-scale outdoor reconstruction underpins city-level digital twins, autonomous-driving simulation and geographic visualization. Panoramic capture brings the data cost down, but block-wise parallel training, the trick that lets 3DGS scale, has not worked for panoramas. PanoLOG opens that road. It shows the frustum is not the only basis for partitioning: render-loss gradients can also carve panoramas into meaningful blocks, restoring parallel block-wise training while saving both memory and model size.

The accompanying Pano360 fills a gap. Until now there was no public benchmark for large-scale panoramic outdoor reconstruction, so follow-up work had no shared target.

This is incremental method engineering, not a new representation paradigm. Its value is narrower: letting a cheaper capture method use a cheaper training method.

Limitations

The authors name two. The framework assumes static scenes; pedestrians and vehicles leave ghosting artifacts and must be masked out first. And Stage I global coarse training still needs a sizable chunk of GPU memory for very large initializations. Partitioning saves Stage II; the Stage I wall still stands.

One more concern on reading: the several thresholds in G²PS (the triangulation range factor, τgrad=0.8, the 10× sky-sphere radius) read as empirical choices, with no systematic sensitivity analysis. Whether they transfer to other scenes is left open. The handheld scenes also start from a much lower baseline (PSNR in the low 20s), so the absolute quality there is still weak; the gain looks good, but whether it clears directly usable is worth judging yourself.

Terms

Source

What people are saying

Related papers

All paper explainers