# VANTAGE-Bench — `data/`

Brief overview of the dataset structure and per-task prompts.
Ground-truth answers are held server-side; only the **question side** of
each annotation ships here.

## Layout

```
data/
├── 2dbbox/                       # 2D bounding-box detection
│   ├── prompt.json
│   └── <sequence>/images/*.jpg
├── dense_captioning/             # Dense video captioning
│   ├── prompt.json
│   └── *.mp4
├── event_verification/           # Binary event classification
│   └── filtered/
│       ├── metropolis_event_verification/{*.mp4, test_annotation.json}
│       ├── tailgating/{location_a, location_b}/{*.mp4, test_annotation.json}
│       └── warehouse_near_miss/{*.mp4, test_annotations.json}
├── pointing/                     # 2D spatial pointing
│   └── Vantage2DPointing.tsv
├── referring/                    # 2D referring expressions
│   └── refdrone_test_llava.json
├── temporal_localization/        # Temporal grounding
│   ├── *.mp4
│   └── data_jsons/annotations/*.json
├── tracking/                     # Stateless single-object tracking
│   └── sot_benchmark.jsonl
└── vqa/                          # Video question answering
    ├── *.mp4
    └── data_jsons/annotations/*.json
```

## Per-task prompts

Tasks without a per-entry `question` field carry a top-level
`prompt.json` with the model instruction (schema: `{"prompt": "<text>"}`).

### `2dbbox/` — 2D Detection
> Locate every instance that belongs to the following categories: `person`. For each instance of the class, report bbox coordinates in JSON format. Do not group instances and report only individual instances. Avoid reporting duplicate instances.

### `dense_captioning/` — Dense Video Captioning
> Describe the notable events in the provided video. Provide the result in json format with `mm:ss.ff` format for time depiction for each event. Use keywords `start`, `end` and `caption` in the json output.

### `vqa/` — Video Question Answering
Per-entry questions in `vqa/data_jsons/annotations/*.json`. Each entry has `{q_uid, question, options, …}`; answer keys (`gt`, `gt_option`, `*_update_*`, etc.) are removed.

### `temporal_localization/` — Temporal Grounding
Per-entry questions in `temporal_localization/data_jsons/annotations/*.json`. Each entry has `{vid, question_id, question, duration, …}`; the `answer` timestamps are removed.

### `event_verification/` — Binary Event Verification
All four annotation files share a unified schema:
`{"bcq": [{id, video, system_prompt, question}, …]}`. The binary `answer`
is removed.

### `pointing/` — 2D Spatial Pointing
`Vantage2DPointing.tsv` — tab-separated. Each row carries the question and multiple-choice options; the last two ground-truth columns are dropped.

### `referring/` — 2D Referring Expressions
`refdrone_test_llava.json` — list of LLaVA conversation entries. Only the `human` turn (the question) is retained; the `gpt` turn (predicted bboxes) and GT meta fields are removed.

### `tracking/` — Stateless Single-Object Tracking
`sot_benchmark.jsonl` — one JSON object per clip with `seq_id`, `scene`, `camera`, `init_bbox` (the seed bounding box you're given as input), `init_frame_id`, and `canonical_frame_ids` (the frames you must predict at). Ground-truth trajectories are held server-side.

## Submitting predictions

See the top-level `README.md` for the eval-server instructions per task.