Wang et al. (2022) — ARXIV-2212.05231
Key takeaways
- Cuts NeuS surface reconstruction from about 8 hours per static object to minutes, and to as little as 20 seconds per frame for a moving-object sequence, at comparable quality
- The speedup is an enabling condition, not an optimisation, and the paper says so: 8 hours per object “makes it almost impossible to apply them to dynamic scenes with thousands of frames”. A thousand frames at 8 hours each is not slow, it is impossible; at 20 seconds it is an afternoon
- Per-frame reconstruction of a sequence warm-starts from the previous frame, which is where the 20-seconds figure comes from and what it assumes
Standing in the anthology
Read — NOTE-046.
A source of SOTA-205 — the fourth paper in the convergence, and the one that carries it out of radiance fields into implicit surfaces. It inherits the hash-encoding line LIT-064 established, which is what makes it corroboration rather than a fifth independent idea.
Worth recording because the record’s own efficiency practices — SOTA-086,
SOTA-019, and the systems work in LIT-102 — are all stated as ratios, and
a ratio does not say whether a result crossed a threshold that changes what can
be attempted. A 2× speedup and a 1000× speedup are qualitatively different
claims, and the record has no way of saying so.