mipal/iPhone360-4dgs360

iPhone360 Dataset - 4dgs360 preprocessed version

iPhone360 is a benchmark dataset for 360° reconstruction of dynamic objects from monocular video, introduced in the paper:

4DGS360: 360° Gaussian Reconstruction of Dynamic Objects from a Single Video Jae Won Jang, Yeonjin Chang, Wonsik Shin, Juhwan Cho, Nojun Kwak Project Page · arXiv

Dataset Description

iPhone360 features real-world dynamic scenes captured with an iPhone, where test cameras are positioned at significantly different angles from training views. This enables evaluation of 360° reconstruction capabilities that existing datasets cannot provide.

Dataset Versions

This dataset is distributed in two versions:

• iPhone360-4dgs360 (this folder) — includes all preprocessing outputs required to reproduce 4DGS360 training and evaluation end-to-end (2D/3D tracks, track-anything masks, refined depth/tracks from AnchorTAPIP3D, cached scene-normalization tensors, etc.). Large footprint.
• iPhone360 — the same RGB/depth/mask/camera/points/splits data, with the 4DGS360-specific intermediate preprocessing outputs above excluded. Much smaller download.

If you're quickly adapting iPhone360 to a new paper/method, we recommend starting with iPhone360 version and evaluating on it first, rather than downloading the full iPhone360-4dgs360.

Scenes

Scene	Description
block2	Dynamic object scene
goat	Dynamic object scene
jacket	Dynamic object scene
jelly	Dynamic object scene
pull-up	Dynamic object scene
walk-around	Dynamic object scene

Data Structure

Each scene contains:

• rgb/ — RGB frames
• depth/ — Depth maps
• masks/ — Object masks
• camera/ — Camera parameters
• splits/ — Train/test split definitions
• points.npy — Initial point cloud
• dataset.json / scene.json / metadata.json — Scene metadata
• flow3d_preprocessed/ — Preprocessed optical flow data
• video_depth_anything/ — Video depth estimates

Citation

If you use this dataset, please cite:

@article{jang2025_4dgs360,
  title     = {4DGS360: 360° Gaussian Reconstruction of Dynamic Objects from a Single Video},
  author    = {Jang, Jae Won and Chang, Yeonjin and Shin, Wonsik and Cho, Juhwan and Kwak, Nojun},
  journal   = {arXiv preprint arXiv:2603.21618},
  year      = {2025},
  url       = {https://arxiv.org/abs/2603.21618}
}