PSVU-Bench
PSVU-Bench is the benchmark for Personalized Streaming Video Understanding (PSVU), introduced in PEARL: Personalized Streaming Video Understanding with Runtime-Defined Concepts.
In PSVU a model watches a video stream, a user defines new concepts (characters, objects or actions) at any timestamp, and later asks questions about these concepts in the present scene or about their past. Every question refers to an exact timestamp of the ongoing stream, so the model must answer causally from what it has seen so far.
- Code: https://github.com/Yuanhong-Zheng/PEARL
- Training set: PSVU-10K
- Mirror: ModelScope
Statistics
| Videos | Concept-Definition | Real-Time QA | Past-Time QA | Total annotations | |
|---|---|---|---|---|---|
Entity split (frame-level/) |
112 | 418 | 922 | 394 | 1,734 |
Action split (video-level/) |
20 | 80 | 359 | 0 | 439 |
| PSVU-Bench | 132 | 498 | 1,281 | 394 | 2,173 |
Each Real-Time and Past-Time question has exactly one category.
| Real-Time category | Questions | Past-Time category | Questions | |
|---|---|---|---|---|
| Presence | 311 | Event-based | 193 | |
| Behavior | 184 | Time-based | 201 | |
| Appearance | 184 | |||
| Location | 134 | |||
| Relation | 109 | |||
| Action | 359 |
Concept-Definition items introduce concepts and are not scored.
Structure
PSVU-Bench/
README.md
LICENSE
SHA256SUMS # checksums of the archive files
frame-level/
annotations/ # 112 JSON files, one per entity video
frame-level.tar.gz.part00 ... # frame-level/videos/ and frame-level/output_clips/
video-level/
annotations/ # 20 JSON files, one per synthetic Action video
video-level.tar.gz # video-level/videos/ and video-level/output_clips/
Download the repository into the data/ folder of the PEARL code and extract the archives there:
huggingface-cli download zyh200727/PSVU-Bench --repo-type dataset --local-dir data
cd data
cat frame-level.tar.gz.part* | tar -xzf -
tar -xzf video-level.tar.gz
The archives are split only to stay below the Hugging Face file-size limit. cat frame-level.tar.gz.part* > frame-level.tar.gz gives the complete archive.
After extraction each split contains:
annotations/: one JSON file per video.videos/: the full source videos (*.mp4).output_clips/<video>/: the video pre-segmented into scene clips (<video>_scene_NNN.mp4) and<video>_clips_info.json, which lists every clip with itsstart_time,end_timeanddurationin seconds.
The entity archive also holds four videos that no annotation refers to (apartment5_2, nezha, qitan, wulimi). They are kept as in the earlier release and are not part of the benchmark.
Annotation format
Each annotation file is a list with one object per video:
[
{
"video_path": "data/frame-level/videos/london.mp4",
"timestamps": [
{"id": 0, "qa_type": "concept definition", "time": "00:00:15",
"question": "This character is called {Carmyn}. Please remember this name. ...",
"gt": "", "concept": true, "concept_name": "{Carmyn}"},
{"id": 2, "qa_type": "current-time qa", "category": "Behavior", "time": "00:04:05",
"question": "Is {Carmyn} lying in bed or leaning against the headboard right now?",
"gt": "D", "concept_name": "{Carmyn}",
"optionA": "None", "optionB": "lying", "optionC": "None", "optionD": "leaning"},
{"id": 4, "qa_type": "past-time qa", "category": "Event-based", "time": "00:07:52",
"question": "When {Carmyn} and {Jeshua} parted ways, was {Carmyn} wearing a scarf?",
"gt": "C", "concept_name": "{Carmyn}, {Jeshua}", "evidence_time": "00:06:06",
"optionA": "None", "optionB": "yes", "optionC": "no", "optionD": "None"}
]
}
]
| Field | Description |
|---|---|
video_path |
Path of the source video relative to the PEARL code root. |
id |
Item index within the video. Items are listed in increasing id. |
qa_type |
concept definition, current-time qa (Real-Time QA) or past-time qa (Past-Time QA). |
category |
Sub-category of a Real-Time or Past-Time question (see below). Concept-Definition items have no category. |
time |
Entity split: stream timestamp (HH:MM:SS) at which the item is issued. |
start_time, end_time |
Action split: the clip (HH:MM:SS) that shows the defined action, or the current clip a question asks about. The question is issued at end_time. |
question |
Concept-definition instruction or question text. Concept names appear in braces, e.g. {Carmyn}. |
optionA ... optionD |
The four answer options. None marks an unused option slot. |
gt |
Letter of the correct option. Empty for Concept-Definition items. |
concept_name |
Concept(s) the item defines or asks about. |
concept |
true for Concept-Definition items. |
evidence_time |
Past-Time QA: timestamp of the historical evidence. |
A few entity items omit concept_name or carry an extra evidence_time. They are kept exactly as annotated.
Categories
Real-Time QA asks about the current scene:
- Presence: whether the concept is present in the current scene ("Is {ConceptName} here now?").
- Behavior: the current action or state of the concept ("What is {ConceptName} doing now?").
- Appearance: transient visual details such as current clothing ("What color is {ConceptName} wearing now?").
- Location: the position of the concept in the scene ("Where is {ConceptName} located in this scene now?").
- Relation: interaction or relationship with other entities or objects ("Who is standing next to {ConceptName} now?").
- Action: whether a user-defined action concept is being performed in the current clip ("Is the person doing {ConceptName} now?").
Past-Time QA asks about earlier content and requires retrieving historical evidence:
- Event-based: the evidence can be localized from a specific event or action ("What was {ConceptName} holding when he was cooking?").
- Time-based: the evidence requires understanding the temporal order of events ("What did {ConceptName} do right before he left the room?").
The original annotations did not record sub-categories. The authors assigned each Real-Time and Past-Time question to one sub-category after annotation, following the definitions above.
Evaluation protocol
Each question is asked four times with the options rotated so that the correct answer appears once in every position. A question counts as correct only if all four rotations are answered correctly. The accuracy of a category is the percentage of its questions answered correctly. The evaluation code is in the PEARL repository.
Sources
- Entity split: videos collected from YouTube and manually filtered for high dynamics, recurring concepts and a resolution of at least 480p. They cover anime, movies and reality shows. Original concept names are replaced with common names sampled from the U.S. SSA baby-names database.
- Action split: synthetic videos whose character models and motions were generated with GPT-6 Astra, then manually arranged and annotated. Each distinct action is a user-defined action concept.
- Annotations: human annotators chose the query timestamps and wrote the questions and answers. Gemini 3.1 Pro Preview generated the multiple-choice distractors for Real-Time and Past-Time QA from the human-specified questions, answers and video clips. Reviewers checked every item against the video.
Intended use
PSVU-Bench is intended for non-commercial research on personalized and streaming video understanding, in particular for evaluating how models register user-defined concepts during a stream and answer questions about them later. It is an evaluation benchmark and should not be used as training data for models that are then evaluated on it. It is not intended for identifying real people or for any surveillance application.
License
The annotations, the metadata and the rendered synthetic Action videos created by the authors are released under CC BY-NC 4.0 (see LICENSE).
The entity source videos and the clips cut from them are third-party content collected from YouTube. They are not licensed by the authors, remain the property of their respective owners and are subject to their original terms. They are provided only to support non-commercial research use of the benchmark.
Privacy and removal requests
The entity videos are publicly available movies, TV series, anime and reality shows and may show identifiable people. The annotations refer to people only through randomly assigned names and describe on-screen content. They contain no personal information about the people shown.
If you own content included in PSVU-Bench, appear in it, or have any other concern, email zyh2022200727@gmail.com with the video file name and the reason for the request. We will review the request and remove the affected videos, clips and annotations from future versions of the dataset.
Citation
@article{zheng2026pearl,
title={PEARL: Personalized Streaming Video Understanding with Runtime-Defined Concepts},
author={Zheng, Yuanhong and An, Ruichuan and Lin, Xiaopeng and Liu, Yuxing and Yang, Sihan and Zhang, Huanyu and Li, Haodong and Zhang, Qintong and Zhang, Renrui and Li, Guopeng and Zhang, Yifan and Li, Yuheng and Zhang, Wentao},
journal={arXiv preprint arXiv:2603.20422},
year={2026},
archivePrefix={arXiv},
eprint={2603.20422},
primaryClass={cs.CV}
}
- Downloads last month
- -