LIFT-Vista

LIFT-Vista is a dataset with large camera viewpoint changes and joint camera-layout annotations.

Overview

	camera.csv	camlayout.csv
Clips	120,898	58,272 (a subset of the camera clips)
Annotations	camera trajectory, caption	camera trajectory, caption, last-frame layout, per-frame box tracks

Every clip has 81 frames at 16 fps (about 5 s) at the native resolution of its source video (98% are 1280×720), encoded with H.264. The release is about 413 GB, of which 407 GB are videos.

Download

Videos and annotations are stored as tar shards in tars/ (about 4 GB per video shard). Extracting them in the dataset root restores the videos/ and annotations/ directories that the CSV paths point to (413 GB in total):

hf download Overdog/LIFT-Vista --repo-type dataset --local-dir LIFT-Vista
cd LIFT-Vista
for f in tars/*.tar; do tar -xf "$f"; done

camlayout.csv only needs the camlayout shards (179 GB):

hf download Overdog/LIFT-Vista --repo-type dataset --local-dir LIFT-Vista --exclude "tars/*camera_only*"
cd LIFT-Vista
for f in tars/*.tar; do tar -xf "$f"; done

Structure

Layout of the dataset root after download and extraction:

LIFT-Vista/
├── camera.csv
├── camlayout.csv
├── videos/<source>/<video_id>/<window>.mp4
└── annotations/<source>/<video_id>/<window>/
    ├── camera_da3.npz              # every clip
    ├── caption.txt                 # every clip
    ├── layout_lastframe.json       # camlayout clips only
    └── layout_track_<W>x<H>.json   # camlayout clips only

<source> is SpatialVID, Sekai or RealEstate10K, <video_id> is the id of the video in the source dataset, and <window> is the window index w of the 81-frame window extracted from the source video.

CSV columns

Column	Description
UID	unique clip id, <video_id>_<window>
resolution	clip resolution, WxH
Video_Path	the clip, relative to the dataset root
Annotation_Path	the annotation directory, relative to the dataset root
CameraFile	camera file inside Annotation_Path
layout_file	last-frame layout inside Annotation_Path (camlayout.csv only)
sam3_track_file	box tracks inside Annotation_Path (camlayout.csv only)
sam3_track_resolution	pixel grid WxH of the track boxes; equal to resolution except for 183 clips whose tracks were computed at 640×352 or 832×480 (rescale x and y separately)
caption	text caption, identical to caption.txt

Camera (camera_da3.npz)

• extrinsic: (81, 3, 4) world-to-camera matrices [R | t], OpenCV convention (x right, y down, z forward).
• intrinsic: (81, 3, 3) pinhole intrinsics in pixels of the clip resolution.

Last-frame layout (layout_lastframe.json)

{
  "instances": [
    {
      "id": 0,
      "category": "chair",
      "bbox": [0, 515, 332, 719],
      "caption": "white upholstered chair with an orange throw pillow"
    },
    ...
  ]
}

Salient foreground objects of the last frame (frame 80), with boxes [x1, y1, x2, y2] in pixels of the clip resolution.

Box tracks (layout_track_<W>x<H>.json)

{
  "source_layout_file": "layout_lastframe.json",
  "track_resolution": "1280x720",               # pixel grid of every box in this file
  "prompt_frame_idx": 80,
  "num_frames": 81,
  "tracks": {
    "0": {"id": 0, "category": "chair", "caption": "...",
          "bboxes": {"37": [x1, y1, x2, y2], ..., "80": [x1, y1, x2, y2]}},   # no box in frames 0..36
    ...
  },
  "frames": {"37": [{"id": 0, "category": "chair", "caption": "...", "bbox": [x1, y1, x2, y2]}, ...], ...}
}

Statistics

camera.csv has 120,898 clips for camera-control training (Stage 1 in the paper) and camlayout.csv has a subset of 58,272 clips with layouts (Stages 2 and 3). The layout annotations contain 5.5 objects per clip on average and cover 8,302 category labels, including indoor objects (chair, window, lamp, cabinet, sofa, ...) and outdoor objects (person, building, car, boat, tree, sign, ...). 84.8% of the layout clips contain objects that are not visible in the first frame and only appear in later views.

(a, b) Distributions of the FoV expansion ratio and the translation distance over all candidate windows before filtering (blue) and ours (red); filtering removes the mass of near-static windows and shifts both metrics toward larger viewpoint changes. (c) Joint distribution of the two camera-motion metrics, with 50% and 90% mass contours per source dataset. (d) Joint distribution of the two content change ratios; the two directions are correlated and complementary but not redundant. (e) Number of training clips per source dataset.

FoV expansion ratio. We uniformly sample $KK$ keyframes from each clip $(K=8)(K\; =\; 8)$. For the $ii$-th keyframe, let ${\mathrm{\Omega }}_{k_i}\subseteq {\mathbb{S}}^2\backslash Omega\_\{k\_i\}\; \backslash subseteq\; \backslash mathbb\{S\}^2$ denote the set of visible viewing directions in a common world coordinate system; its spherical area is estimated using uniformly sampled directions on the unit sphere. The FoV expansion ratio

$$r_{\mathrm{F}\mathrm{o}\mathrm{V}}=\frac{\mid \bigcup _{i=1}^K{\mathrm{\Omega }}_{k_i}\mid }{\mid {\mathrm{\Omega }}_{k_1}\mid }r\_\{\backslash mathrm\{FoV\}\}\; =\; \backslash frac\{\backslash left|\; \backslash bigcup\_\{i=1\}^\{K\}\; \backslash Omega\_\{k\_i\}\; \backslash right|\}\{\backslash left|\; \backslash Omega\_\{k\_1\}\; \backslash right|\}$$

measures the total viewing region covered by the clip relative to the first frame.

Translation distance. The accumulated camera translation is

$$d_{\mathrm{t}\mathrm{r}\mathrm{a}\mathrm{n}\mathrm{s}}=\sum _{i=1}^{K-1}{\parallel {\mathbf{o}}_{k_{i+1}}-{\mathbf{o}}_{k_i}\parallel }_2,d\_\{\backslash mathrm\{trans\}\}\; =\; \backslash sum\_\{i=1\}^\{K-1\}\; \backslash left\backslash |\; \backslash mathbf\{o\}\_\{k\_\{i+1\}\}\; -\; \backslash mathbf\{o\}\_\{k\_i\}\; \backslash right\backslash |\_2\; ,$$

where ${\mathbf{o}}_{k_i}\backslash mathbf\{o\}\_\{k\_i\}$ denotes the camera origin of the $ii$-th keyframe.

Content change ratios (CCR). Since camera motion alone does not directly measure changes in visible scene content, we additionally compute a patch-level CCR between the first and last frames using DINOv2. Let $\{{\mathbf{p}}_i{\}}_{i=1}^N\backslash \{\backslash mathbf\{p\}\_i\backslash \}\_\{i=1\}^\{N\}$ and $\{{\mathbf{q}}_j{\}}_{j=1}^N\backslash \{\backslash mathbf\{q\}\_j\backslash \}\_\{j=1\}^\{N\}$ denote their ${\mathrm{\ell }}_2\backslash ell\_2$-normalized patch embeddings. The last-frame CCR is

where $\mathbb{I}[\cdot ]\backslash mathbb\{I\}[\backslash cdot]$ denotes the indicator function, $⟨\cdot ,\cdot ⟩\backslash langle\; \backslash cdot,\; \backslash cdot\; \backslash rangle$ denotes the cosine similarity, and $\tau \backslash tau$ is a similarity threshold $(\tau =0.5)(\backslash tau\; =\; 0.5)$. $r_{lf\text{-}CCR}r\_\{lf\backslash text\{-\}CCR\}$ measures the fraction of last-frame patches unmatched in the first frame (new content); $r_{ff\text{-}CCR}r\_\{ff\backslash text\{-\}CCR\}$ is computed analogously in the reverse direction to measure content leaving the initial view (lost content).

Usage

import json, os
import numpy as np
import pandas as pd

root = "LIFT-Vista"
df = pd.read_csv(os.path.join(root, "camlayout.csv"))
row = df.iloc[0]

video_path = os.path.join(root, row.Video_Path)            # 81 frames, 16 fps
ann_dir = os.path.join(root, row.Annotation_Path)

cam = np.load(os.path.join(ann_dir, row.CameraFile))
w2c, K = cam["extrinsic"], cam["intrinsic"]                # (81, 3, 4), (81, 3, 3)

with open(os.path.join(ann_dir, row.sam3_track_file)) as f:
    tracks = json.load(f)["tracks"]
tw, th = map(int, row.sam3_track_resolution.split("x"))   # pixel grid of the boxes
for obj in tracks.values():
    frames = sorted(int(k) for k in obj["bboxes"])
    print(obj["category"], "| visible in frames", frames[0], "to", frames[-1], "|", obj["caption"])

The LIFT training code reads these CSVs directly; see the code repository.

Citation

Coming soon.

Please also cite the source datasets:

@misc{wang2025spatialvidlargescalevideodataset,
      title={SpatialVID: A Large-Scale Video Dataset with Spatial Annotations}, 
      author={Jiahao Wang and Yufeng Yuan and Rujie Zheng and Youtian Lin and Jian Gao and Lin-Zhuo Chen and Yajie Bao and Yi Zhang and Chang Zeng and Yanxi Zhou and Xiaoxiao Long and Hao Zhu and Zhaoxiang Zhang and Xun Cao and Yao Yao},
      year={2025},
      eprint={2509.09676},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2509.09676}, 
}

@article{li2025sekai,
      title={Sekai: A Video Dataset towards World Exploration}, 
      author={Zhen Li and Chuanhao Li and Xiaofeng Mao and Shaoheng Lin and Ming Li and Shitian Zhao and Zhaopan Xu and Xinyue Li and Yukang Feng and Jianwen Sun and Zizhen Li and Fanrui Zhang and Jiaxin Ai and Zhixiang Wang and Yuwei Wu and Tong He and Jiangmiao Pang and Yu Qiao and Yunde Jia and Kaipeng Zhang},
      journal={arXiv preprint arXiv:2506.15675},
      year={2025}
}

@article{zhou2018stereo,
  author    = {Zhou, Tinghui and Tucker, Richard and Flynn, John and Fyffe, Graham and Snavely, Noah},
  title     = {Stereo Magnification: Learning View Synthesis using Multiplane Images},
  journal   = {ACM Transactions on Graphics (SIGGRAPH)},
  volume    = {37},
  number    = {4},
  articleno = {65},
  year      = {2018},
  url       = {https://dl.acm.org/doi/10.1145/3197517.3201323}
}