README.md

---
license: cc-by-4.0
language:
  - en
pretty_name: VISTA
task_categories:
  - visual-question-answering
  - text-generation
tags:
  - video
  - audio
  - timeseries
  - 3d
  - egocentric
  - multimodal
  - accessibility
size_categories:
  - "n<1K"
configs:
  - config_name: default
    data_files:
      - split: train
        path: train.csv
      - split: validation
        path: val.csv
      - split: test
        path: test.csv
---

# VISTA: A Multimodal Egocentric Benchmark for Goal-Oriented Perception and Assistance to Blind and Low-Vision Users

### Dataset Summary

VISTA (**Visually Impaired Scene and Task Annotations**) is a multimodal egocentric benchmark for goal-oriented perception and assistance to blind and low-vision users. It contains **997 first-person video clips** spanning ten assistive task categories, with **1,063 scene descriptions, 1,191 question–answer pairs, and 2,067 goal–instruction pairs**.

The dataset pairs RGB video with SLAM-camera and eye-camera recordings, inertial and environmental measurements, gaze, hand-tracking, and trajectory signals. Spoken task requests accompany the annotations. Clips are organized into **497 training, 50 validation, and 450 test examples**, with no clip overlap between splits.

### Supported Tasks

- **Scene description:** describe objects, layout, obstacles, and hazards in the first-person view.
- **Visual question answering:** answer questions about the observed scene.
- **Goal-conditioned guidance:** provide instructions for achieving an assistive goal.

### Languages

English.

## Quick Start

The default Hub configuration exposes the annotation CSVs as train, validation, and test tables. Each row represents one clip; media and sensor files are located using its `clip_id`.

For an already downloaded dataset:

```python
import csv
from pathlib import Path

root = Path("/path/to/VISTA")
with (root / "test.csv").open(encoding="utf-8-sig", newline="") as handle:
    sample = next(csv.DictReader(handle))

clip_id = sample["clip_id"]
clip = root / "test" / clip_id.split("_")[0] / clip_id
rgb_video = clip / f"{clip_id}.mp4"
slam_left = clip / "vrs" / "slam_left.mp4"
question_audio = clip / "generated-audio" / "q1.wav"
```

Use the question or goal index corresponding to the requested annotation when selecting its WAV.

## Splits and annotations

| Split | Clips | Scene-description annotations | Question–answer pairs | Goal–instruction pairs | Generated task WAVs |
|---|---:|---:|---:|---:|---:|
| Train | 497 | 555 | 622 | 1,091 | 2,210 |
| Validation | 50 | 52 | 59 | 108 | 217 |
| Test | 450 | 456 | 510 | 868 | 1,828 |
| **Total** | **997** | **1,063** | **1,191** | **2,067** | **4,255** |

Clip IDs are disjoint across splits. Counts are based on the supplied CSVs. A clip can have multiple reference annotations, so annotation totals are not counts of independent videos. Each clip has one scene-request WAV, regardless of its number of scene-description references.

## Task categories

| ID | Category | Clips |
|---|---|---:|
| TC01 | Object localization | 100 |
| TC02 | Hazard detection | 99 |
| TC03 | Scene / situational awareness | 102 |
| TC04 | Navigation | 101 |
| TC05 | Social / person description, including facial-expression tasks | 97 |
| TC06 | Distance estimation | 99 |
| TC07 | Tool use / task instructions | 100 |
| TC08 | Object identification | 100 |
| TC09 | Reading / text recognition | 102 |
| TC10 | Motion / dynamic scene understanding | 97 |

## File structure

The dataset repository root contains the three annotation CSVs and their corresponding media folders:

```text
VISTA/
├── README.md
├── train.csv
├── val.csv
├── test.csv
├── train/
│   └── TC01/
│       └── TC01_002/
│           ├── TC01_002.mp4
│           ├── generated-audio/
│           │   ├── q1.wav
│           │   ├── g1.wav
│           │   └── scene-discription-prompt.wav
│           ├── vrs/
│           │   ├── slam_left.mp4
│           │   ├── slam_right.mp4
│           │   ├── eye.mp4
│           │   ├── imu_left.csv
│           │   ├── imu_right.csv
│           │   ├── magnetometer.csv
│           │   ├── barometer.csv
│           │   ├── rgb_timestamps.csv
│           │   └── time_domain_mapping.csv
│           └── mps/
│               ├── eye_gaze/
│               │   └── general_eye_gaze.csv
│               ├── hand_tracking/
│               │   └── hand_tracking_results.csv
│               └── slam/
│                   ├── closed_loop_trajectory.csv
│                   ├── open_loop_trajectory.csv
│                   ├── online_calibration.jsonl
│                   ├── semidense_points.csv.gz
│                   └── semidense_observations.csv.gz
├── val/
│   └── TC01/ ... TC10/
└── test/
    └── TC01/ ... TC10/
```

This is a representative layout, not a complete listing of supplementary metadata files. The number of `qN.wav` and `gN.wav` files varies with the annotations. 

## Annotation schema

Each CSV row represents one clip, identified by `clip_id`.

- `scene_description_N`: a reference scene description.
- `question_N` and `answer_N`: a question and its paired reference answer.
- `goal_N` and `instruction_N`: a goal and its paired reference instruction.

Empty cells represent absent annotations. Preserve the original indices when matching annotations to audio; do not renumber populated columns after dropping empty cells.

## Known Limitations

1. **Geographic coverage:** Recordings were collected in Boston, so the dataset may not represent environments and assistive scenarios in other locations.
2. **Missing tracking data:** Hand-tracking results are missing for 15 clips, and derived SLAM data (trajectories, point clouds, observations, and calibration) for 11 different clips. In total, 26 clips lack at least one derived modality. RGB, camera videos, gaze, inertial/environmental measurements, and generated task audio are available for all 997 clips.
3. **Safety-critical use:** Benchmark performance does not establish that a model is safe for real-world navigation or assistance. Deployment requires evaluation with blind and low-vision users.


## Licensing Information

The dataset is licensed under **Creative Commons Attribution 4.0 International (CC BY 4.0)**. See the [license terms](https://creativecommons.org/licenses/by/4.0/). The accompanying benchmark code has a separate MIT license.