Does dynamic reconstruction work at minute-scale capture in the wild?

The dynamic-reconstruction line in this corpus reports almost all of its quantitative results on short clips of a single deformable subject in relatively controlled conditions. Nerfies (Park et al., 2020) is the canonical example: ~10–20 second selfie videos, indoor lighting, one moving subject. HyperNeRF (Park et al., 2021) extends the deformation field to handle topological change but still operates on the same scale of input. Dynamic View Synthesis from Dynamic Monocular Video (Gao et al., 2021) is similar.

The 2023–2024 splatting analogues — 4D Gaussian Splatting (Wu et al., 2023) and 4D-Rotor Gaussian Splatting (Duan et al., 2024) — solve the rendering-cost half of the problem convincingly. Both report real-time playback of dynamic novel views, and both use the same explicit-primitive trick that 3DGS used to displace NeRF in the static case. But the captures they evaluate on are still short. K-Planes (Fridovich-Keil et al., 2023) is the most aggressive on duration — its planar factorisation makes the memory growth linear in time rather than quadratic — but its public results are still on minutes-or-less captures.

What "in the wild" means here is concrete: a hand-held monocular phone video, several minutes long, of a scene with multiple moving subjects and non-trivial illumination change (a busy street, a marketplace). None of the papers in this corpus benchmarks on that regime. The likely failure modes are easy to enumerate: the deformation MLP's capacity is fixed at training time; the canonical-frame assumption breaks when no single canonical state covers the whole capture; the per-Gaussian temporal trajectories in 4DGS-class methods grow linearly in primitive count, which the densification schedule wasn't designed for.

What would resolve this: a benchmark dataset of multi-minute monocular dynamic captures with calibrated ground-truth, plus a head-to-head evaluation of K-Planes, 4DGS, 4D-Rotor-GS, HyperNeRF, and Nerfies on it. Bonus points for ablating capture length explicitly so the failure modes have a name. Until that exists, the field is publishing real-time 4D rendering on captures that are short enough for the failure modes to hide.