README.md

---
language: en
task_categories:
- keypoint-detection
tags:
- 3d-tracking
- multi-view
- point-cloud
- computer-vision
- robotics
- synthetic-data
- real-world-data
- pytorch
- pytorch-hub
license: other
---

# Multi-View 3D Point Tracking Datasets

This repository hosts the training and evaluation datasets associated with the paper [**Multi-View 3D Point Tracking**](https://huggingface.co/papers/2508.21060).

**Project Page:** [https://ethz-vlg.github.io/mvtracker/](https://ethz-vlg.github.io/mvtracker/)
**Code/Github Repository:** [https://github.com/ethz-vlg/mvtracker](https://github.com/ethz-vlg/mvtracker)

## Abstract

We introduce the first data-driven multi-view 3D point tracker, designed to track arbitrary points in dynamic scenes using multiple camera views. Unlike existing monocular trackers, which struggle with depth ambiguities and occlusion, or prior multi-camera methods that require over 20 cameras and tedious per-sequence optimization, our feed-forward model directly predicts 3D correspondences using a practical number of cameras (e.g., four), enabling robust and accurate online tracking. Given known camera poses and either sensor-based or estimated multi-view depth, our tracker fuses multi-view features into a unified point cloud and applies k-nearest-neighbors correlation alongside a transformer-based update to reliably estimate long-range 3D correspondences, even under occlusion. We train on 5K synthetic multi-view Kubric sequences and evaluate on two real-world benchmarks: Panoptic Studio and DexYCB, achieving median trajectory errors of 3.1 cm and 2.0 cm, respectively. Our method generalizes well to diverse camera setups of 1-8 views with varying vantage points and video lengths of 24-150 frames. By releasing our tracker alongside training and evaluation datasets, we aim to set a new standard for multi-view 3D tracking research and provide a practical tool for real-world applications.

## Dataset Details

To benchmark multi-view 3D point tracking, we provide preprocessed versions of three datasets:

-   **MV-Kubric**: a synthetic training dataset adapted from single-view Kubric into a multi-view setting.
-   **Panoptic Studio**: evaluation benchmark with real-world activities such as basketball, juggling, and toy play (10 sequences).
-   **DexYCB**: evaluation benchmark with real-world hand–object interactions (10 sequences).

You can download and extract them as (~72 GB after extraction):

```bash
# MV-Kubric (simulated + DUSt3R depths)
wget https://huggingface.co/datasets/ethz-vlg/mv3dpt-datasets/resolve/main/kubric-multiview--test.tar.gz -P datasets/
wget https://huggingface.co/datasets/ethz-vlg/mv3dpt-datasets/resolve/main/kubric-multiview--test--dust3r-depth.tar.gz -P datasets/
tar -xvzf datasets/kubric-multiview--test.tar.gz -C datasets/
tar -xvzf datasets/kubric-multiview--test--dust3r-depth.tar.gz -C datasets/
rm datasets/kubric-multiview*.tar.gz

# Panoptic Studio (optimization-based depth from Dynamic3DGS)
wget https://huggingface.co/datasets/ethz-vlg/mv3dpt-datasets/resolve/main/panoptic-multiview.tar.gz -P datasets/
tar -xvzf datasets/panoptic-multiview.tar.gz -C datasets/
rm datasets/panoptic-multiview.tar.gz

# DexYCB (Kinect + DUSt3R depths)
wget https://huggingface.co/datasets/ethz-vlg/mv3dpt-datasets/resolve/main/dex-ycb-multiview.tar.gz -P datasets/
wget https://huggingface.co/datasets/ethz-vlg/mv3dpt-datasets/resolve/main/dex-ycb-multiview--dust3r-depth.tar.gz -P datasets/
tar -xvzf datasets/dex-ycb-multiview.tar.gz -C datasets/
tar -xvzf datasets/dex-ycb-multiview--dust3r-depth.tar.gz -C datasets/
rm datasets/dex-ycb-multiview*.tar.gz
```

For licensing and usage terms, please refer to the original datasets from which these preprocessed versions are derived.

## Sample Usage

This dataset repository contains the data for the MVTracker model. With minimal dependencies in place (as described in the [GitHub repository](https://github.com/ethz-vlg/mvtracker#quick-start)), you can try MVTracker directly via **PyTorch Hub**:

```python
import torch
import numpy as np
from huggingface_hub import hf_hub_download

device = "cuda" if torch.cuda.is_available() else "cpu"
mvtracker = torch.hub.load("ethz-vlg/mvtracker", "mvtracker", pretrained=True, device=device)

# Example input from demo sample (downloaded automatically)
sample = np.load(hf_hub_download("ethz-vlg/mvtracker", "data_sample.npz"))
rgbs = torch.from_numpy(sample["rgbs"]).float()
depths = torch.from_numpy(sample["depths"]).float()
intrs = torch.from_numpy(sample["intrs"]).float()
extrs = torch.from_numpy(sample["extrs"]).float()
query_points = torch.from_numpy(sample["query_points"]).float()

with torch.no_grad():
    results = mvtracker(
        rgbs=rgbs[None].to(device) / 255.0,
        depths=depths[None].to(device),
        intrs=intrs[None].to(device),
        extrs=extrs[None].to(device),
        query_points_3d=query_points[None].to(device),
    )

pred_tracks = results["traj_e"].cpu()  # [T,N,3]
pred_vis = results["vis_e"].cpu()      # [T,N]
print(pred_tracks.shape, pred_vis.shape)
```

## Citation

If you find our repository useful, please consider giving it a star ⭐ and citing our work:

```bibtex
@inproceedings{rajic2025mvtracker,
  title     = {Multi-View 3D Point Tracking},
  author    = {Raji{\v{c}}, Frano and Xu, Haofei and Mihajlovic, Marko and Li, Siyuan and Demir, Irem and G{\"u}ndo{\u{g}}du, Emircan and Ke, Lei and Prokudin, Sergey and Pollefeys, Marc and Tang, Siyu},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year      = {2025}
}
```