Gaussian streaming now budgets anchors per object: OCLDA-GS beats IGS by 0.169 dB on N3DV and cuts inter-keyframe PSNR drop by 9.2%
Synopsis
Online Gaussian streaming previously spread anchors over the whole scene and propagated motion by neighborhood interpolation, so motion from one object could leak across object boundaries. OCLDA-GS groups Gaussians by object: optical flow shifts a fixed anchor budget toward fast-moving objects, interpolation stays inside each group, and a per-object rigid-motion estimate conditions a learned per-anchor local deformation correction. Under matched refinement budgets it improves PSNR over Instant Gaussian Stream by 0.169 dB on N3DV, 0.202 dB on Meeting Room and 0.199 dB on Google Immersive, while reducing the mean PSNR drop between keyframes by 9.2% and 13.8%.
Interpretation
Online streaming Gaussian reconstruction now allocates motion anchors per object rather than across the whole scene, and motion propagation near object boundaries is confined to the same object group. Earlier methods such as IGS and ReConGS assumed neighboring Gaussians share one 3D deformation trajectory and picked anchors by farthest point or hierarchical sampling, which under a limited anchor budget can give excessive support to stationary regions. On the three dynamic multi-view datasets N3DV, Meeting Room and Google Immersive, with the same initialization and refinement budget as IGS, OCLDA-GS reaches 33.046 dB, 30.290 dB and 23.798 dB PSNR, respectively 0.169 dB, 0.202 dB and 0.199 dB above IGS.
Each object first gets one overall rigid transform, and small per-anchor deformation corrections are learned on top of it, so non-rigid detail such as cloth wrinkles need not be recovered by per-frame refinement. Earlier methods only interpolated anchor motion features and did not separate an object's shared rigid component from local non-rigid deformation inside the object. The rigid transform is a least-squares fit by the Kabsch algorithm to each object's predicted anchor correspondences, and the residual network takes frozen optical-flow features, anchor motion features, positional encoding, rigid-relative displacement, semantic features and temporal offset as input; on N3DV and Meeting Room the mean PSNR drop from frames 0 to 4 after a keyframe falls from 0.622 to 0.565 dB and from 1.927 to 1.662 dB.
Perspective
Evaluation covers the three dynamic multi-view video datasets N3DV, Meeting Room and Google Immersive, with IGS, ReConGS, 3DGStream and QUEEN as online comparisons and offline methods included only as context; reconstruction time excludes initial Gaussian reconstruction.
The method depends on object grouping quality, and inaccurate grouping can limit object-aware motion propagation; the rigid-motion mechanism assumes each object group has a useful dominant rigid transform, which may be inaccurate for articulated objects; the learned local deformation correction updates Gaussian positions only and does not correct orientation, scale or appearance.
