pretty_name: PanoVideoBench
language:
- en
task_categories:
- visual-question-answering
size_categories:
- 1K<n<10K
tags:
- panoramic-video
- video-question-answering
- multimodal
- reasoning
configs:
- config_name: default
data_files:
- split: test
path: PanoVideoBench.jsonl
PanoVideoBench
Think in 360°: Benchmarking and Bootstrapping Panoramic Video Reasoning
Summary
A panoramic video records the complete surrounding scene, while the evidence needed to answer a question may appear in different directions and at different times. PanoVideoBench evaluates panoramic video understanding through 1,000 manually annotated four-choice questions over 223 real panoramic videos from 360+x. The questions cover appearance, spatial relations, actions, motion, and temporal events.
Each question includes a semantic category and question type, together with diagnostic labels describing how the relevant view is found and which panoramic capabilities are needed to answer. These labels support evaluation beyond a single overall score.
Benchmark Results
Results from the paper under the default evaluation protocol. All scores are choice accuracy (%), with higher values indicating better performance. Bold marks the best result in each column.
| Model | Overall | APP | SPA | ACT | MOT | TEM |
|---|---|---|---|---|---|---|
| Gemini 3.7 Flash | 53.1 | 65.7 | 49.5 | 52.4 | 40.9 | 36.5 |
| GPT-5.6-sol | 49.6 | 59.9 | 46.0 | 51.9 | 33.5 | 41.9 |
| Gemma-4-31B-it | 47.0 | 58.4 | 48.5 | 44.6 | 34.8 | 27.0 |
| Nemotron Omni 30B-A3B | 46.3 | 59.9 | 46.0 | 41.6 | 31.1 | 35.1 |
| InternVL3.5-8B | 44.9 | 53.8 | 36.1 | 49.4 | 34.1 | 39.2 |
| Qwen3.8-27B | 44.8 | 54.7 | 44.1 | 46.8 | 31.1 | 27.0 |
| Qwen3.5-9B | 43.2 | 55.4 | 36.6 | 42.5 | 35.4 | 27.0 |
| Eagle2.5-8B | 39.6 | 50.5 | 35.6 | 38.6 | 32.3 | 21.6 |
| MiniCPM-V-4.6 | 39.0 | 48.9 | 37.1 | 34.3 | 34.1 | 25.7 |
| GLM-4.1V-9B-Thinking | 38.0 | 46.5 | 38.6 | 32.2 | 31.1 | 32.4 |
| PanoWorld-Hstar | 42.9 | 51.7 | 45.0 | 36.5 | 34.1 | 37.8 |
| PanoVideoReasoner | 55.7 | 59.6 | 52.5 | 60.1 | 48.8 | 48.6 |
APP, SPA, ACT, MOT, and TEM denote appearance and detail perception, spatial and viewpoint reasoning, action and state understanding, motion and identity tracking, and spatio-temporal event reasoning, respectively.
Benchmark Statistics
| Property | Value |
|---|---|
| Questions | 1,000 |
| Videos | 223 |
| Total video duration | 23.06 hours |
| Mean video duration | 372 seconds |
| Semantic categories | 5 |
| Question types | 15 |
| Choices per question | 4 |
| Category | Scope | Questions |
|---|---|---|
| APP | Panoramic Appearance and Detail Perception | 327 |
| SPA | Spherical Spatial and Viewpoint Reasoning | 202 |
| ACT | Omnidirectional Action and State Understanding | 233 |
| MOT | Cross-View Motion and Identity Tracking | 164 |
| TEM | Panoramic Spatio-Temporal Event Reasoning | 74 |
Question Types
| Category | Types |
|---|---|
| APP | APP-SCN: Scene and Layout Understanding; APP-ENT: Entity and Attribute Recognition; APP-FGR: Fine-Grained Visual Recognition |
| SPA | SPA-OBS: Observer-Centric Spatial Relations; SPA-OBJ: Object-Centric Spatial Relations; SPA-ORI: Orientation and Viewpoint Reasoning |
| ACT | ACT-INT: Action and Interaction Recognition; ACT-STA: State and Transition Understanding; ACT-MUL: Multi-Entity Activity Understanding |
| MOT | MOT-TRK: Trajectory and Identity Tracking; MOT-COM: Camera–Object Motion Disambiguation |
| TEM | TEM-LOC: Event Localization; TEM-ORD: Event Ordering; TEM-COUNT: Event Counting; TEM-AGG: Cross-Time Evidence Aggregation |
Diagnostic Annotations
Each question has one target-search label:
| Label | Name | Questions |
|---|---|---|
| S0 | Direction-Guided Search | 312 |
| S1 | Landmark-Guided Search | 337 |
| S2 | Event-/Moving-Entity-Guided Search | 265 |
| S3 | Unconstrained Panoramic Search | 86 |
Capability labels describe the operations needed after finding the relevant evidence:
| Label | Name | Questions |
|---|---|---|
| C0 | Find, then Read | 343 |
| C1 | See the Fine Detail | 109 |
| C2 | Reason on the Sphere | 349 |
| C3 | Track Across Views | 239 |
| C4 | Integrate Evidence Across Sphere–Time | 137 |
C0 identifies questions resolved by local readout. C1–C4 can co-occur, so their counts are not mutually exclusive.
Data Format
PanoVideoBench.jsonl contains one question per line. The release contains annotations and video identifiers; video files are obtained separately from the source collection.
| Field | Description |
|---|---|
id |
Unique question ID. |
video_id |
Source video identifier. |
question |
English question. |
options |
Four English choices keyed by A, B, C, and D. |
correct_option |
Correct choice letter. |
answer |
Text of the correct choice. |
category |
Primary category code: APP, SPA, ACT, MOT, or TEM. |
question_type |
Secondary type code. |
search |
One target-search label, S0–S3. |
capabilities |
List of capability labels, C0–C4. |
Diagnostic labels are provided for all 1,000 questions.
Evaluation Protocol
Use the complete ERP video as the visual input. Uniformly sample frames over the full duration at a target rate of 1 FPS, with at most 128 frames, and resize the long edge to 1,280 pixels. The default setting uses the original ERP representation without perspective mosaics or orientation overlays. The paper uses each model's default generation temperature and a maximum of 32,768 output tokens, including thinking tokens.
Answers and diagnostic labels are evaluation annotations, not model inputs.
Evaluation Metrics
Compare the model's final selected option with correct_option. Accuracy is 100 × number of correct answers / number of evaluated questions.
- Overall: accuracy over all 1,000 questions. Each question has equal weight; this is not an unweighted average of the five category scores.
- APP / SPA / ACT / MOT / TEM: accuracy within each semantic category, using 327 / 202 / 233 / 164 / 74 questions, respectively.
- Diagnostic accuracy: accuracy on the subset carrying each search label (S0–S3) or capability label (C0–C4). A question with multiple capability labels contributes to each corresponding subset.
Citation