# 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
│   ├── *.mp4 (under videos/)
│   └── data_jsons/annotations/*.json
├── pointing/                     # 2D spatial pointing
│   └── VANTAGE_2DPointing.jsonl
├── 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
carries exactly three fields, scoped to model inference:
- `q_uid` — video/sample identifier; resolves against `vqa/videos/`
- `question` — natural-language question text
- `options` — list of MCQ answer choices used to build the prompt

Ground-truth (`gt_option`, `answer`) and per-question metadata
(`industry`, `event_type`, `task_type`, `dimension`, `start_time`,
`end_time`, `video_duration`) are not included in the public
annotations.

### `temporal_localization/` — Temporal Grounding
Per-entry questions in `temporal_localization/data_jsons/annotations/*.json`.
Each entry carries exactly three fields, scoped to model inference:
- `vid` — video identifier; resolves against `temporal_localization/`
- `question_id` — stable annotation identifier (reproducibility key)
- `question` — temporal-localization query

Ground-truth timestamps and per-question metadata (`industry`,
`event_type`, `task_type`, `duration`) are not included in the public
annotations.

### `event_verification/` — Binary Event Verification
Per-entry questions in `event_verification/data_jsons/annotations/*.json`
(four files: `VANTAGE_EventVerification.json` — 67 entries,
`tailgating_location_a.json` — 28, `tailgating_location_b.json` — 22,
`warehouse_near_miss.json` — 46; 163 total). Each file is a top-level
list of sample objects with schema
`[{id, video, system_prompt, question}, …]` — matching the
`vqa/` and `temporal_localization/` annotation layout — where `video`
is the basename (e.g. `example.mp4`) and `id` is the stem
(e.g. `example`), resolving against `event_verification/videos/`. The
binary `answer` is removed.

### `pointing/` — 2D Spatial Pointing
`VANTAGE_2DPointing.jsonl` — one JSON object per line, 1,005 lines,
8 fields: `index, question_id, image_path, question, A, B, C, D`. Each
line carries the question and four multiple-choice options (`A`–`D`);
each option is an `x,y` pair (string `"x,y"`) in the **normalized
`0–1000` coordinate system** (both components in `[0, 1000]` relative
to the image dimensions). `index` is an integer in `[0, 1004]`.
Ground-truth fields (`answer`, `target_point`) are held server-side and
are not included in the public JSONL.

### `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.