An arXiv preprint reports 59% lower per-frame optimization time and 85% lower storage cost for S²GS than for QUEEN in tests on an RTX 4090 GPU. S²GS is designed for free-viewpoint video, in which reconstructed scenes can be rendered from changing angles.
The evaluation tested three versions—S²GS-full, S²GS-fast and S²GS-edge—on benchmark scenes, Jetson AGX Orin hardware and a physical testbed.
The gains come with a quality trade-off
The benchmark evaluation included six N3DV scenes, three Meet Room scenes and three 300-frame ENeRF Outdoor sequences involving six actors.
On N3DV, the average row for S²GS-full reported 32.76 dB PSNR and 0.952 SSIM, two image-quality scores. The same row listed 101k Gaussian primitives—the model’s 3D scene elements—0.11 MB of storage, 4.12 seconds of training and 482 frames per second for rendering.
The scores were not uniformly higher across measures. On ENeRF Outdoor, S²GS-full reported 25.94 dB PSNR versus QUEEN at 26.13 dB, while SSIM was 0.863 versus 0.843 and LPIPS was 0.124 versus 0.154. QUEEN led on PSNR, while S²GS-full led on SSIM and LPIPS in that comparison.
Sparse by design
At its core, S²GS uses a streaming octree, a tree-like map of 3D space, with structured hierarchical gating. It combines differentiable Gumbel-Sigmoid sampling, multi-level straight-through discretization and sparse regularization for residual updates.
In a component ablation, the multi-level straight-through method reported 32.76 dB PSNR, 0.952 SSIM and 679 active gates, while a binary straight-through version reported 29.85 dB, 0.934 SSIM and two active gates.
A separate ablation combining the structure and sparsity components reported 32.46 dB PSNR, 64k Gaussians, 401 active gates, 0.08 MB of storage, 1.96 seconds of training and 513 frames per second. Without either component, the reported values were 31.69 dB, 129k Gaussians, 129k active gates, 0.67 MB, 2.86 seconds and 383 frames per second.
Across five evaluated datasets, active Gaussian ratios ranged from 1.99% to 12.94%, while active gates represented 9.90% to 23.94% of active Gaussians. The pattern persisted from 1/2 to 1/16 input resolution.
An edge-device test
On a Jetson AGX Orin, S²GS-edge reported 33.26 dB PSNR, 0.968 SSIM and 0.036 LPIPS. It reported 8.622 seconds of training, 67.94 frames per second of rendering, 11.18 watts of power and 96.39 joules per frame, compared with QUEEN-s at 32.90 dB, 17.04 seconds, 52.06 frames per second, 14.08 watts and 239.92 joules per frame.
The physical testbed used nine industrial cameras and three USB 3.0 hubs. In the Install RAM scene, S²GS-edge used 88k Gaussian primitives and 0.26 MB of storage, with 4.60 seconds of per-frame optimization and 59.37 frames per second of rendering. The authors describe this setup as proof-of-concept evidence rather than a long-term field deployment.
Where the method strains
The fixed hierarchy is the main constraint described in the paper. Later frames were less reliable when they required new spatial support, with reported failures under occlusion or disocclusion, large motion outside the supported area and fine-scale stochastic dynamics.
Reported quality and efficiency metrics were averaged over three independent runs. For self-measured baseline comparisons, the authors matched data, views, camera parameters, initialization and hardware; values imported from original papers were reference values only.
Paper data and sources
Original title: S$^2$GS: Structured Sparse Gaussian Streaming for Efficient Free-Viewpoint Video Reconstruction on Edge-IoT Devices
Authors: Yiwei Li, Jiannong Cao, Weixun Gao et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text