dynadust3r-unofficial-weights

DynaDUSt3R (Unofficial) — Model Weights

This repository hosts PyTorch .pth checkpoint files for the DynaDUSt3R model — an external, unofficial reimplementation of the model proposed in the paper Stereo4D: Learning How Things Move in 3D from Internet Stereo Videos by Linyi Jin, Richard Tucker, Zhengqi Li, David Fouhey, Noah Snavely, and Aleksander Hołyński.

Overview

DynaDUSt3R extends DUSt3R to dynamic scenes by predicting per-pixel 3D points and 3D motion trajectories from stereo video sequences.
These weights were obtained from an independent reimplementation and are provided solely for research and reproducibility purposes.

The accompanying codebase can be found in the dynadust3r-unofficial repository.
Instructions for running inference are available in the inference example script.

Contents

• .pth checkpoint files corresponding to trained DynaDUSt3R models.
• Optional SHA256SUMS.txt for integrity verification.

These checkpoints are compatible with the model definitions in the dynadust3r-unofficial repository.

Usage

Please follow the setup and inference steps described in the main dynadust3r-unofficial repository.
The weights can be loaded using torch.load() as shown in the inference example.

Note: .pth files use Python pickle and should only be loaded from trusted sources.

ADT Evaluation — motion_left_weighted

The table below summarizes motion_left_weighted metrics extracted from evaluation logs using regex.
All runs used num_pairs = 80. Lower EPE3D is better; higher δ[email protected] / δ[email protected] are better.
Sorting: by EPE3D (ascending).

---

References

• Stereo4D Project Page
• DUSt3R Paper (CVPR 2024)

---

License

These weights are provided for research and non-commercial use only.
This repository is unofficial and not affiliated with the authors of Stereo4D.