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.
Annotations and questions for non-commercial research. No third-party video is included; you must obtain the source videos (ACaM, SpatialVID, Blender open films) under their own terms. Access is reviewed manually. 仅供非商业研究；不含任何第三方视频，源视频需按其原始条款自行获取。下载需人工审批。
Log in or Sign Up to review the conditions and access this dataset content.
Long-form Camera Motion Bench / 长视频相机运动评测集
English · 中文
Annotations, question sets, build recipes and model outputs for long-video camera-motion understanding (1–10 min).
No video or frames are included. The stitched videos can be rebuilt exactly from the ACaM clips listed in the recipes.
Companion models and report: Zeteng/longform-camera-motion-models.
Contents
| Path | What |
|---|---|
stitched/pilot164.jsonl |
164 stitched long videos (38–209 s, mean 84 s, about 13 segments). Each segment gives its ACaM test source clip (src), the source dataset, the human label, start / end and the boundary type. Real and synthetic clips are never mixed in one video; adjacent labels differ; a label appears at most twice per video |
stitched/length_set.jsonl |
60 / 30 / 12 videos at 1 / 3 / 10 min (about 11 / 33 / 108 segments), with the ACaM source clip of every segment |
stitched/timeline_sft_{train,val}.jsonl |
training data of Timeline SFT: 1,654 stitched ACaM-train videos; frame paths are relative to ACaM's 32-frame train frames |
questions/length_set_qa_v1.jsonl |
What / When / Order, 4-choice, plus one open timeline item per video. The v1 "When" questions are position-biased; do not use them |
questions/length_set_when_v2_debiased.jsonl |
debiased When questions; the gold's rank by start time is uniform; the "earliest option" prior is 27–31% |
questions/e0_what_at_time.jsonl, e1_when_ordinal_order.jsonl |
earlier E0 / E1 questions on the pilot set (frame indices refer to the rebuilt pilot videos) |
spatialvid_raw_timelines/ |
SpatialVID annotations placed back on the YouTube timeline (69,961 runs ≥ 1 min). Also: priority tiers A / B / C for runs ≥ 3 min (runs_prio.csv) and change statistics (quality.json). Labels are pose-derived and automatic, not human; the footage is mostly vehicle or walking motion |
open_films/{sintel,tears_of_steel,cosmos_laundromat}/prelabel_camsft_auto.json |
shot boundaries (ffmpeg scdet) and CamSFT-8B labels for three CC BY Blender films. Automatic pre-labels, NOT ground truth |
outputs/baselines*/ |
raw outputs of 6 VLMs × 2 visual budgets (64 frames; 64 frames per minute) on the length set |
outputs/two_stage_camsft/ |
CamSFT-8B per segment on true and on Timeline-SFT boundaries |
outputs/mstcn_length_preds/ |
MS-TCN timelines on the length set |
Sources and licences
| Component | Source | Terms |
|---|---|---|
| Stitched benchmark labels, recipes | ACaM test / train (arXiv 2607.03043) | ACaM's terms apply; videos must be obtained from ACaM |
| SpatialVID-RAW timelines | SpatialVID / SpatialVID-RAW annotations | SpatialVID's licence (non-commercial research); YouTube videos are not redistributed |
| Open-film pre-labels | Sintel (CC BY 3.0), Tears of Steel (CC BY 3.0), Cosmos Laundromat (CC BY 4.0), © Blender Foundation, www.blender.org | the films' CC BY licences; download from Blender's official sources |
| Questions, statistics, model outputs | this work | non-commercial research |
Known issues
- Every boundary in the stitched videos is a hard cut.
length_set_qa_v1When questions are biased (see above).- The film labels are automatic.
- The 10-minute tier is small (12 videos).
中文
用于长视频（1–10 分钟）相机运动理解的标注、题目、构建配方和模型输出。不含任何视频或帧；拼接视频可以按配方中列出的 ACaM 片段精确重建。配套模型与报告：Zeteng/longform-camera-motion-models。
内容
| 路径 | 说明 |
|---|---|
stitched/pilot164.jsonl |
164 条拼接长视频（38–209 秒，平均 84 秒，约 13 段）。每段给出 ACaM 测试集来源片段（src）、来源数据集、人工标签、起止时间和边界类型。真实与合成片段不混在一条视频里；相邻标签不同；同一标签在一条视频中最多 2 次 |
stitched/length_set.jsonl |
1 / 3 / 10 分钟各 60 / 30 / 12 条（约 11 / 33 / 108 段），每段附 ACaM 来源片段 |
stitched/timeline_sft_{train,val}.jsonl |
Timeline SFT 的训练数据：1,654 条 ACaM 训练片段拼接的长视频 |
questions/length_set_qa_v1.jsonl |
What / When / Order 选择题，以及每条视频一道开放的时间线题。v1 的 When 题有位置偏差，请勿使用 |
questions/length_set_when_v2_debiased.jsonl |
去偏后的 When 题，正确答案的位置均匀分布，“总选最早”只有 27–31% |
questions/e0_*、e1_* |
早期 pilot 集上的 E0 / E1 题目 |
spatialvid_raw_timelines/ |
SpatialVID 标注放回 YouTube 时间轴后的长片段（≥ 1 分钟的有 69,961 段），另含 ≥ 3 分钟片段的 A / B / C 分档与变化统计。标签由位姿自动生成，未经人工核验，且以行车、步行等载体运动为主 |
open_films/*/prelabel_camsft_auto.json |
三部 CC BY Blender 电影的镜头边界与 CamSFT 自动标签。是预标注，不是真值 |
outputs/* |
6 个 VLM × 两种视觉预算、CamSFT 逐段识别、MS-TCN 的原始输出 |
来源与许可
- ACaM： 适用 ACaM 的条款，视频需从 ACaM 获取。
- SpatialVID： 适用 SpatialVID 的许可（非商业研究），YouTube 视频不再分发。
- Blender 电影： CC BY 3.0 / 4.0，© Blender Foundation | www.blender.org，请从官方渠道下载。
- 本工作产出的部分（题目、统计、模型输出）： 仅供非商业研究。
已知问题
- 拼接视频的边界都是硬切镜。
- v1 的 When 题有偏差。
- 电影标签是自动生成的。
- 10 分钟档样本少（12 条）。
- Downloads last month
- -