# Dynamic Point Maps: A Versatile Representation for Dynamic 3D Reconstruction

Source: [Dynamic Point Maps: A Versatile Representation for Dynamic 3D Reconstruction](<https://www.sota2.com/research/paper/68fa77811e209131cea73a41>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2025-10-26T04:20:40.032Z

Authors: Edgar Sucar, Zihang Lai, Eldar Insafutdinov, Andrea Vedaldi

Published: 2025-03-20T16:41:50.000Z

## Abstract

DUSt3R has recently shown that one can reduce many tasks in multi-view geometry, including estimating camera intrinsics and extrinsics, reconstructing the scene in 3D, and establishing image correspondences, to the prediction of a pair of viewpoint-invariant point maps, i.e., pixel-aligned point clouds defined in a common reference frame. This formulation is elegant and powerful, but unable to tackle dynamic scenes. To address this challenge, we introduce the concept of Dynamic Point Maps (DPM), extending standard point maps to support 4D tasks such as motion segmentation, scene flow estimation, 3D object tracking, and 2D correspondence. Our key intuition is that, when time is introduced, there are several possible spatial and time references that can be used to define the point maps. We identify a minimal subset of such combinations that can be regressed by a network to solve the sub tasks mentioned above. We train a DPM predictor on a mixture of synthetic and real data and evaluate it across diverse benchmarks for video depth prediction, dynamic point cloud reconstruction, 3D scene flow and object pose tracking, achieving state-of-the-art performance. Code, models and additional results are available at https://www.robots.ox.ac.uk/~vgg/research/dynamic-point-maps/.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [4D Reconstruction on Kubric-F](<https://www.sota2.com/research/sota/4d-reconstruction-on-kubric-f.md>): 4D Reconstruction; Kubric-F; EPE P0 (t1): 0.033.

- [4D Reconstruction on Kubric G](<https://www.sota2.com/research/sota/4d-reconstruction-on-kubric-g.md>): 4D Reconstruction; Kubric G; EPE P0 (t1): 0.04.

- [Dense Tracking on Waymo](<https://www.sota2.com/research/sota/dense-tracking-on-waymo.md>): Dense Tracking; Waymo; EPE: 0.103.

- [Dense Tracking on PointOdyssey](<https://www.sota2.com/research/sota/dense-tracking-on-pointodyssey.md>): Dense Tracking; PointOdyssey; EPE: 0.114.

- [Dense Tracking on Kubric-F](<https://www.sota2.com/research/sota/dense-tracking-on-kubric-f.md>): Dense Tracking; Kubric-F; EPE: 0.088.

- [Dense Tracking on Kubric G](<https://www.sota2.com/research/sota/dense-tracking-on-kubric-g.md>): Dense Tracking; Kubric G; EPE: 0.109.

- [4D reconstruction on PointOdyssey](<https://www.sota2.com/research/sota/4d-reconstruction-on-pointodyssey.md>): 4D Reconstruction; PointOdyssey; EPE P0(t1): 0.103.

- [4D reconstruction on Waymo](<https://www.sota2.com/research/sota/4d-reconstruction-on-waymo.md>): 4D Reconstruction; Waymo; EPE P0 (t=1): 0.083.

- [Dense 3D Tracking on PointOdyssey 10-frame snippets](<https://www.sota2.com/research/sota/dense-3d-tracking-on-pointodyssey-10-frame-snippets.md>): Dense 3D Tracking; PointOdyssey 10-frame snippets; EPE: 0.114.

- [Dense 3D Tracking on Kubric-F (10-frame snippets)](<https://www.sota2.com/research/sota/dense-3d-tracking-on-kubric-f-10-frame-snippets.md>): Dense 3D Tracking; Kubric-F (10-frame snippets); Endpoint Error (EPE): 0.088.

- [Dense 3D Tracking on Kubric-G 10-frame snippets](<https://www.sota2.com/research/sota/dense-3d-tracking-on-kubric-g-10-frame-snippets.md>): Dense 3D Tracking; Kubric-G 10-frame snippets; Endpoint Error (EPE): 0.109.

- [Dense 3D Tracking on Waymo (10-frame snippets)](<https://www.sota2.com/research/sota/dense-3d-tracking-on-waymo-10-frame-snippets.md>): Dense 3D Tracking; Waymo (10-frame snippets); EPE: 0.103.

- [Two-view tracking on Kubric-F (test)](<https://www.sota2.com/research/sota/two-view-tracking-on-kubric-f-test.md>): Two-view tracking; Kubric-F (test); Tracking Error P0 (t1): 0.05.

- [Two-view tracking on Kubric-G (test)](<https://www.sota2.com/research/sota/two-view-tracking-on-kubric-g-test.md>): Two-view tracking; Kubric-G (test); P0 (t1): 0.068.

- [Two-view tracking on PointOdyssey (test)](<https://www.sota2.com/research/sota/two-view-tracking-on-pointodyssey-test.md>): Two-view tracking; PointOdyssey (test); Tracking Error P0 (t1): 0.103.

- [Two-view tracking on Waymo (test)](<https://www.sota2.com/research/sota/two-view-tracking-on-waymo-test.md>): Two-view tracking; Waymo (test); Error P0 (t1): 0.085.

- [Video Depth Estimation on Bonn](<https://www.sota2.com/research/sota/video-depth-estimation-on-bonn.md>): Video Depth Estimation; BONN; Relative Error (Rel): 0.064.

- [Camera Pose Estimation on TUM dynamics](<https://www.sota2.com/research/sota/camera-pose-estimation-on-tum-dynamics.md>): Camera pose estimation; TUM dynamics; ATE: 0.056.

- [Video Depth Estimation on Sintel](<https://www.sota2.com/research/sota/video-depth-estimation-on-sintel.md>): Video Depth Estimation; Sintel; Delta Threshold Accuracy (1.25): 58.

- [Camera Pose Estimation on TUM-dynamic](<https://www.sota2.com/research/sota/camera-pose-estimation-on-tum-dynamic.md>): Camera pose estimation; TUM-dynamic; ATE: 0.056.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2503.16318>)

[PDF](<https://arxiv.org/pdf/2503.16318.pdf>)
