Long-range 4D Motion Reconstruction

Minsik Jeon, Jay Karhade, Deva Ramanan, Shubham Tulsiani
Carnegie Mellon University  ·  Equal advising

Point4D chains motion reconstruction across overlapping video chunks by re-querying each predicted 3D endpoint directly in the next chunk — no reprojection, no matching — enabling robust 4D tracking over 300+ frames.

0:00 / 0:00
Abstract

We introduce Point4D, a feed-forward model for 4D reconstruction of long-range video sequences. Point4D is able to reliably infer dense per-point 3D trajectories across multi-hundred video sequences, unlike existing 4D methods that are limited to short input windows of at most a few dozen frames. A key innovation that enables this is our flexible 3D query-based motion decoder that decouples trajectory prediction from image-plane visibility. The predicted 3D endpoints are then directly re-queried in the next chunk without re-projection or matching. Furthermore, we show that extracting and reusing a visual descriptor from an arbitrary frame where the point is visible leads to superior performance than purely relying on the source patch. Overall, Point4D achieves state-of-the-art performance across diverse long-video tracking benchmarks spanning over 200 frames and substantially outperforms previous feed-forward 4D methods.

Method: Training

0:00 / 0:00

Results

Long-video 4D tracking

On ~200-frame sequences (chunk 48 / overlap 8), Point4D outperforms every feed-forward baseline across datasets and visibility conditions.

Long-video tracking metrics

Iterative trackers (SpaTrackerV2, TAPIP3D) are compared in the paper — Point4D outperforms or matches their accuracy while running an order of magnitude faster. Full tables and additional datasets in the paper.

Long-horizon stability

Per-chunk trajectory error as the video progresses — Point4D's (red plot) error grows slowest across chunk boundaries.

Per-chunk EPE/APD on Dynamic Replica

^ Dynamic Replica (200 frames)

Per-chunk EPE/APD on Panoptic Studio

^ Panoptic Studio (150 frames)

Per-chunk EPE (↓) and APD (↑); chunk 48 / overlap 8.

Single chunk tracking

Point4D performs comparably within a single chunk (up to 64 frames) — its gains come from chaining.

Method LSFOdysseyDyn. ReplicaPStudio
EPE↓APD↑EPE↓APD↑EPE↓APD↑
2DDA3 + CoTracker3 0.500.690.110.890.330.63
Iter.TAPIP3D 0.350.660.870.500.300.65
SpatialTrackerV2 0.340.680.690.620.210.75
FFSt4RTrack 0.560.480.170.810.410.53
Trace Anything 0.760.420.340.660.250.72
Any4D 0.270.720.070.930.280.66
V-DPM 0.140.850.140.840.170.79
4RC 0.160.870.070.950.290.66
Point4D (Ours) 0.270.730.090.910.230.73

EPE (↓) and APD (↑); 1st 2nd 3rd.