Ost et al. (2020) — ARXIV-2011.10379
Key takeaways
- The first neural rendering method to decompose a dynamic scene into a scene graph — nodes carrying object transformation and radiance — so novel arrangements can be rendered, not merely novel views of a fixed configuration. Learned “only by observing a video of this scene”
- Rendering a new viewpoint and rendering a new arrangement are different capabilities, and the second requires objects to be separable entities. Encoding a whole scene into one network makes the first easy and the second impossible
- One shared implicit function plus a per-object latent describes all objects, which is what allows unseen objects at unseen poses to be placed
- Demonstrated on synthetic and real automotive data
Standing in the anthology
Read — NOTE-035. The document is Superseded and the reading does not argue
with that.
The durable statement: a capability that requires manipulating parts cannot be
obtained from a representation without parts — an argument about representation
design rather than capacity. It is the same argument LIT-070’s attribute-binding
failure makes from the other side, where a representation lacking the needed
structure could not carry it no matter how good the decoder was.
“Temporal coherence” is not a claim this paper makes; the dynamics come from object transformations in a graph, not from a coherence constraint over time.