Point4D: Long-range 4D Motion Reconstruction

What it claims

Point4D is a feed-forward model that infers dense per-point 3D trajectories across multi-hundred-frame videos — benchmarks spanning over 200 frames — where existing 4D methods are capped at short windows of "at most a few dozen frames." The abstract claims state-of-the-art across diverse long-video tracking benchmarks and says it "largely outperforms" the previous feed-forward 4D method (unnamed here). This is the substance the summary could not carry: the specific gap it targets (long-range, 200+ frames) and that it is feed-forward rather than per-scene optimization.

How it works

Three design choices are named:

• A 3D query-based motion decoder that decouples trajectory prediction from image-plane visibility. Trajectories are predicted in 3D rather than being tied to where the point is visible in the image.
• Chunked re-querying without re-projection or matching: predicted 3D endpoints from one chunk are directly re-queried as inputs in the next chunk. This is how it chains across hundreds of frames without the usual re-projection/matching step between windows.
• Visual descriptor reuse: extracting a descriptor from an arbitrary frame where the point is visible outperforms relying solely on the source patch. The abstract presents this as an empirical finding, implying an ablation exists in the paper.

flowchart LR
  A[Video chunk N] --> B[3D query decoder]
  B --> C[3D trajectory endpoints]
  C -->|re-query, no re-projection| B2[chunk N+1]
  D[descriptor from any visible frame] --> B

Limits

This is an abstract only; no metric values, benchmark names, model size, compute, code, or licence are on this page, though a project page is referenced ("this https URL"). The specific benchmarks, the number the SOTA claim rests on, and the identity of the "previous feed-forward 4D method" it outperforms are not stated — the reader must open the paper or project page for any number. Note the submission date (8 Sep 2026) and v1 status: unreviewed preprint. Whether the descriptor-reuse and decoupled-visibility design hold up outside the tested long-video tracking setting is not addressed.