ARROW

Arbitrary Reconstruction and Tracking of 4D Observations in the Wild

Ilya Fradlin    Christian Schmidt    Jens Piekenbrinck
Karim Knaebel    Gonzalo Martin Garcia    Bastian Leibe

RWTH Aachen University

Overview

ARROW is a feed-forward model that unifies 3D reconstruction and 3D point tracking from arbitrary image sets. Its point-query decoder supports observations from different viewpoints and capture times, enabling reconstruction and tracking from moving-camera videos, multiple video streams, and unordered photo collections.

Usage

The snippet below briefly illustrates point queries and task helpers. For a quick start and more details, see GitHub.

import torch
from arrow import ARROW, Queries
from arrow.queries import pixel_grid
from arrow.tasks.reconstruction import reconstruction_queries
from arrow.tasks.tracking import tracking_queries

model = ARROW.from_pretrained("ifradlin/arrow", device="cuda")
images = torch.rand(1, 3, 3, 378, 504, device="cuda")  # RGB in [0, 1]
H, W = images.shape[-2:]

with torch.inference_mode():
    # Option 1: point queries.
    queries = Queries(pixel_grid(H, W, stride=4),
                      source_frame=0, target_frame=1, reference_frame=0)
    predictions = model(images, queries)

    # Option 2: task helpers, sharing one encoding.
    encoded = model.encode(images)
    reconstruction = model.decode(encoded, reconstruction_queries(images))
    tracks = model.decode(encoded, tracking_queries(images, stride=16))

Weights are released under CC BY-NC 4.0 for non-commercial use.

Citation

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

@article{fradlin2026arrow,
    title         = {{ARROW}: Arbitrary Reconstruction and Tracking of 4D Observations in the Wild},
    author        = {Fradlin, Ilya and Schmidt, Christian and Piekenbrinck, Jens and Knaebel, Karim and Martin Garcia, Gonzalo and Leibe, Bastian},
    year          = 2026,
    journal       = {arXiv preprint arXiv:2610.01314},
}