You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

OakInk2 Ego→Exo 采样数据集

从 OakInk-v2 全部 627 个序列中采样的同步多视角图像,用于 ego→exo 视角生成(Qwen-Image-Edit-2511 LoRA 训练)。图像来自 OakInk2(Zhan et al., CVPR 2024),按原数据集的 CC BY-SA 4.0 许可再分发,使用时请引用 OakInk2。

规模

序列 时间点 图像对 (ego→exo)
train 564 15,048 45,144
val 63 1,693 5,079
合计 627 16,741 50,223

每个时间点有 4 张同步图(848×480 PNG):egocentric 和 3 个外部相机 allocentric_left / allocentric_right / allocentric_top,共 66,964 张,约 33GB。

文件

  • images/train_NNN.tar, images/val_NNN.tar:图像分片,每片约 2GB,按完整序列打包(一个序列不会跨分片)。解压后是 images/<sequence_id>/<role>/<frame>.png,和 manifest 里的路径一致。
  • shards.json:每个分片包含哪些序列。
  • manifests/pairs_all.jsonl:全部 50,223 个图像对,每行有 source_image(ego)、target_image(exo)、target_view、prompt、primitive_labels、split 等字段。
  • manifests/splits/train_90.jsonl, val_10.jsonl:按序列划分的 train / val。
  • manifests/sequences.csv:每个序列的时间点数、对数、时长、动作 primitive。
  • manifests/sampled_timestamps.csv:每个采样时间点(帧号、时间、boundary/regular、动作)。
  • seq_meta.tar:每个序列一个 json,含相机定义 cam_def、采样参数和每个时间点的相机内外参 calib(取自 OakInk2 标注)。

采样规则

帧号是 120fps mocap 帧号,4 个相机共享。保留所有动作段边界帧;边界之间每 200 帧加一个点(每个边界后重新起算),离下一个边界不到 100 帧的点去掉;每个点对齐到 8 帧以内最近的有图帧。

视角和 prompt

相机 位置 prompt(训练时前面加 turn the view to )
allocentric_left 佩戴者左手边 exocentric view back-right quarter view eye-level shot wide shot
allocentric_right 佩戴者右手边 exocentric view back-left quarter view elevated shot wide shot
allocentric_top 上方 exocentric view back view elevated shot wide shot

下载(国内)

export HF_ENDPOINT=https://hf-mirror.com
hf auth login   # 私有仓库需要 token
hf download wrc1415691977/OakInk2_Ego2Exo_Sampled --repo-type dataset --local-dir OakInk2_Ego2Exo_Sampled
cd OakInk2_Ego2Exo_Sampled && for f in images/*.tar; do tar xf "$f"; done
Downloads last month
41