MyVEBench human-review package

This repository contains a portable human-review package for 104 candidate Source–Real Target pairs covering 104 unique video UUIDs. It is a pre-release candidate pool, not the frozen benchmark test set.

Use the online reviewer

Open https://huggingface.co/spaces/ZenithVoyager/MyVEBench. The interface starts in English. Use the language button in the upper-right corner to switch to Chinese. Enter an optional Reviewer ID, watch both complete videos, complete the two core judgments, and add a note only when useful. The overall routing decision is derived automatically.

Reviews remain in that browser's localStorage; nothing is uploaded automatically. Export JSON regularly. The Reviewer ID is included in the filename and payload. To resume on another browser or computer, open the same candidate release and use Import JSON.

Temporal-length policy

The reviewer plays the original complete Source and Real Target clips. A clip is not rejected merely because it contains fewer than 81 native frames. After human acceptance, fixed-length model/evaluator views are produced by mapping each clip independently over its complete relative time span to 81 frames. This normalization does not assert frame-by-frame Source–Target correspondence, and the Real Target remains one valid action exemplar rather than a unique pixel-level ground truth.

Run the complete reviewer locally

Recommended download method:

python3 -m pip install -U huggingface_hub
hf download ZenithVoyager/MyVEBench --repo-type dataset --local-dir MyVEBench
cd MyVEBench
python3 -m http.server 8000

Then open http://localhost:8000/review.html. Directly opening review.html also works in most browsers, but a local HTTP server gives the most consistent video behavior. Keep review.html and media/ together.

Git users may instead run:

git lfs install
git clone https://huggingface.co/datasets/ZenithVoyager/MyVEBench
cd MyVEBench
python3 -m http.server 8000

Verify the download with:

shasum -a 256 -c SHA256SUMS

Decision protocol

Review every pair on only two core judgments:

1. Instruction–Target alignment: the Target clearly realizes the requested new action, while the relevant subject, objects, and scene remain corresponding to the Source;
2. Camera edit: select yes when the Target introduces a perceptible camera change relative to the Source, and no for an almost-static camera or negligible jitter. When this is yes, one conditional question asks whether the editing instruction explicitly requests the same camera change. An unrequested perceptible camera change is rejected; a requested one is routed to Camera.

There is one independent reject flag for a Target action that works only as a direct continuation of a special state reached at the end of the Source. This does not require the Target to start from the Source's first frame or initial state. The page derives Action, Camera, Reject, or Uncertain automatically. Notes are optional and may be left blank. The Real Target is an action exemplar, not a unique frame-by-frame ground truth.

中文使用说明

打开在线 Space 后，点击右上角 中文。逐条完整观看 Source 与 Real Target，只判断 Instruction–Target 对齐，并直接标记 Target 相对 Source 是否存在 Camera edit；仅在选“是”时，再判断 同一种 camera 变化是否由 editing instruction 明确要求。未要求的可感知运镜直接拒绝。若动作只能作为 Source 末尾特殊状态的直接续接，再勾选独立拒绝原因。这里不要求 Target 从 Source 首帧或初始状态开始。系统自动给出 “Action / Camera / 拒绝 / 不确定”，备注可以留空。页面只在本机浏览器自动保存，不会把结果自动传回服务器；请 定期点击 导出 JSON。换电脑时可用 导入 JSON 恢复同一批候选的记录。

页面播放的是完整原始视频；不能仅因 Source 或 Target 原生不足 81 帧而拒绝样本。人审通过后，模型与部分 evaluator 所需的 81 帧版本会对 Source 和 Target 分别按完整相对时间进度确定性生成，不代表二者逐帧配准。

本地使用时，按上方命令下载整个 Dataset 仓库，在目录中执行 python3 -m http.server 8000，再访问 http://localhost:8000/review.html。不要只复制 HTML，必须同时保留 media/。

Provenance and licensing

The clips are a filtered research subset derived from https://huggingface.co/datasets/MultiEventVideo/MEV. The upstream dataset is marked mixed-third-party-licenses; this repository does not apply a new blanket license to the videos. Users are responsible for following the provenance and license terms recorded by the upstream dataset. See NOTICE.md.

Included files

• review.html: fully portable bilingual reviewer (English by default);
• media/: deduplicated complete Source/Target clips;
• site_manifest.json: sanitized case metadata;
• human_review_template.json and .csv: blank machine-readable templates;
• media_receipt.json, release_receipt.json, and SHA256SUMS: integrity records.