RealEstate10K-Cap

RealEstate10K-Cap pairs 25,000 camera trajectories from RealEstate10K with natural-language descriptions of the camera motion. It is the training data of TKCAM.

split	clips
train	21,250
validation	2,500
test	1,250

Each clip has 119–279 frames (median 190), sampled at about 30 fps from a real-estate YouTube video. Captions are 16–73 words long (median 35).

What is included: camera trajectories, captions, part-of-speech-tagged captions, motion tags, the original RealEstate10K camera poses, intrinsics, frame timestamps and the YouTube URL.

What is not included: video frames. The videos belong to their YouTube owners. scripts/download_frames.py fetches each video and extracts the frames at the exact RealEstate10K timestamps (see Getting the frames).

Quick start

from datasets import load_dataset
import numpy as np

ds = load_dataset("yhz-linearalgebra/RealEstate10K-Cap")          # or load_dataset("path/to/this/folder")
clip = ds["train"][0]
traj = np.asarray(clip["trajectory_12d"])             # (num_frames, 12)
print(clip["caption"])

To train TKCAM directly, export per-clip .npy files in its dataset layout:

python scripts/export_npy.py --data-dir data --out-dir RealEstate10K-Cap_npy --format 12d
# --format 9d writes the [rot6d, xyz] layout used by CLaTr and the TKCAM benchmark

Fields

field	type	description
id	str	dataset id, 000000 … (not contiguous)
split	str	train / val / test (a random split of the 25k clips)
re10k_clip_id	str	RealEstate10K clip id (name of the original camera file)
re10k_official_split	str	RealEstate10K split the clip came from (train: 22,562, test: 2,438)
youtube_url, youtube_video_id	str	source video
num_frames	int	N, number of frames in the clip
timestamps_us	list[int] (N)	frame timestamps in the YouTube video, in microseconds
intrinsics	[fx, fy, cx, cy]	normalized by image width/height; constant within every clip
re10k_pose_w2c	list[N × 12]	original RealEstate10K poses, unchanged: row-major 3×4 world-to-camera, OpenCV axes (+X right, +Y down, +Z forward)
trajectory_12d	list[N × 12]	camera trajectory in TKCAM's 12-D format (see below)
caption	str	camera-motion description
caption_pos	str	the caption as space-separated word/POS tokens (spaCy en_core_web_sm, universal POS tags, punctuation removed)
motion_guidance	str	deterministic motion description computed from trajectory_12d and given to the captioner
motion_tags	list[str]	motion classes detected from trajectory_12d (see below)
in_benchmark_test	bool	clip is one of the RealEstate10K prompts in the TKCAM text-to-camera benchmark

trajectory_12d

Each row is camera-to-world, relative to the first frame, in OpenGL camera axes (+X right, +Y up, −Z forward). Row 0 is the identity pose.

[x, y, z, dx, dy, dz, r11, r21, r31, r12, r22, r32]
 position  velocity      first two columns of the 3×3 rotation (third = col0 × col1; forward = −col2)

Velocity is pos[t] − pos[t−1]; row 0 uses pos[1] − pos[0]. Each of the 12 channels is then smoothed with a Gaussian (σ = 1 frame). The exact computation from re10k_pose_w2c is in scripts/re10k_to_trajectory.py:

C_i = W_i⁻¹ · diag(1, −1, −1, 1)     # world-to-camera (OpenCV) → camera-to-world (OpenGL)
S_i = C_0⁻¹ · C_i                     # relative to frame 0

Running python scripts/re10k_to_trajectory.py --data-dir data recomputes the trajectories from the poses and reports the difference (float32 rounding only).

Frame alignment: row i (0-indexed) of trajectory_12d ↔ timestamps_us[i] ↔ re10k_pose_w2c[i] ↔ frame file frame_{i+1:06d}_{timestamps_us[i]:012d}.jpg.

stats/Mean.npy and stats/Std.npy are the per-channel mean and std of trajectory_12d over all 25k clips. TKCAM uses them for normalization.

motion_tags

The tags are dollies_forward, dollies_backward, tracks_left, tracks_right, pans_left, pans_right, tilts_up, tilts_down, moves_up, moves_down, arcs_left, arcs_right and static. A clip may have several tags or none.

tag	clips	tag	clips
dollies_forward	19,854	tilts_down	2,424
arcs_left	6,518	dollies_backward	2,117
arcs_right	6,218	moves_down	1,526
moves_up	4,639	tracks_right	1,382
pans_right	2,906	tracks_left	1,256
pans_left	2,542	tilts_up	873
static	667

The 25k clips were sampled from 32,412 candidates (see curation_report.json). The sampling balanced motion classes that were computed on an earlier version of the trajectories, so the table above, not the report, is the current distribution.

How the captions were made

Each clip was captioned by Qwen3-VL-4B-Instruct from 32 frames sampled uniformly from the clip, together with motion_guidance. The model was asked to keep the given motion and to add the main scene elements, in 1–2 sentences.

Getting the frames

pip install yt-dlp opencv-python pyarrow
python scripts/download_frames.py --data-dir data --out-dir frames --split test

The script writes frames/<re10k_clip_id>/frame_<k:06d>_<timestamp_us:012d>.jpg, 640×360 by default (--height). This is the layout TKCAM's --frame_dir expects. Clips from the same video share one download. Some videos have been taken down since RealEstate10K was released; the clips that could not be fetched are listed in frames/failed.json.

License

The annotations in this dataset (captions, tags and derived trajectories) are released under CC BY 4.0. The camera poses come from RealEstate10K, which Google LLC releases under CC BY 4.0. The videos belong to their creators and are not redistributed here.

Citation

@misc{yang2026tkcam,
  title={TKCAM: Text and Keyframe to Camera Trajectory Generation},
  author={Yang, Haozhe and Dou, Zhiyang and Gu, Zekai and Lin, Cheng and Wang, Wenping and Liu, Yuan and Komura, Taku},
  year={2026},
  eprint={2610.11105},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2610.11105}
}

@article{zhou2018stereo,
  title   = {Stereo Magnification: Learning View Synthesis using Multiplane Images},
  author  = {Zhou, Tinghui and Tucker, Richard and Flynn, John and Fyffe, Graham and Snavely, Noah},
  journal = {ACM Transactions on Graphics (Proc. SIGGRAPH)},
  volume  = {37},
  number  = {4},
  year    = {2018}
}