Dusting the Phantom: Gaussian Allocation for Fading Dynamic Content

Dusting the Phantom. Gaussian Allocation for Fading Dynamic Content. By Minhyeok Roh and six others, from Yonsei University, the Electronics and Telecommunications Research Institute, and two other institutions. Published on the fifteenth of October. No code repository is linked to it.

Dynamic Gaussian representations have made high-quality, real-time volumetric video increasingly practical. Hard motion can still break the reconstruction, though. Small objects moving fast may vanish, and fine textures in highly dynamic regions are often lost. Training for longer, or adding more Gaussians, may still not bring those small, fast objects back. The obvious remedies, more time and more primitives, do not reliably fix it, and that failure is the clue the paper follows.

The authors' diagnosis is that capacity is not the real bottleneck. What matters is where and when the Gaussian primitives get allocated. Existing allocation strategies work from where the Gaussians currently sit, and they tend to overlook where, in space and time, a relocated Gaussian ends up. In other words, they decide based on the present layout and lose track of the destination. The answer is two complementary strategies, each aimed at one side of that allocation problem: where new Gaussians should start, and where moved ones should go.

The first is Seed Injection. It finds dynamic regions that are under-reconstructed, using changes in appearance over time together with reconstruction errors. It then places new Gaussians at geometrically plausible 3D locations worked out from multi-view observations, independent of where existing Gaussians happen to be.

The second is Temporal Biased Relocation. It uses time-resolved optimization signals to steer where, in space and in time, relocated Gaussians are placed. Both strategies keep the Gaussian budget fixed, and neither changes the underlying representation. Nothing is added to the model's capacity. The same number of primitives is simply put to work in different places, which is the direct test of the claim that allocation, not capacity, was the limit.

Experiments run across multiple dynamic-scene datasets. They show better reconstruction quality, especially in highly dynamic regions, including the recovery of small, rapidly moving objects and of fine-grained appearance details. The abstract does not give the figures.

Phantoms, in this account, come from misplacement rather than scarcity. With the same number of Gaussians and the same representation, sending them to the right regions at the right moments is what brings the fading content back.