README.md

---
dataset_info:
- config_name: Application
  features:
  - name: id
    dtype: large_string
  - name: task
    dtype: large_string
  - name: subtask
    dtype: large_string
  - name: q_type
    dtype: large_string
  - name: question
    dtype: large_string
  - name: answer
    dtype: large_string
  - name: referrable
    dtype: bool
  - name: image
    dtype: image
  - name: image_alt
    dtype: image
  - name: cell
    dtype: large_string
  - name: target_layer
    dtype: large_string
  - name: subdivision
    dtype: large_string
  - name: img_center
    dtype: large_string
  - name: features
    dtype: large_string
  splits:
  - name: train
    num_bytes: 19868297
    num_examples: 561
  - name: test
    num_bytes: 19868297
    num_examples: 561
  download_size: 38753116
  dataset_size: 39736594
- config_name: ENC_Perception
  features:
  - name: id
    dtype: large_string
  - name: task
    dtype: large_string
  - name: subtask
    dtype: large_string
  - name: q_type
    dtype: large_string
  - name: question
    dtype: large_string
  - name: answer
    dtype: large_string
  - name: referrable
    dtype: bool
  - name: visible
    dtype: bool
  - name: image
    dtype: image
  - name: image_alt
    dtype: image
  - name: cell
    dtype: string
  - name: target_layer
    dtype: string
  - name: img_center
    dtype: large_string
  - name: feature
    dtype: string
  splits:
  - name: train
    num_bytes: 338935136
    num_examples: 11850
  - name: test
    num_bytes: 338935136
    num_examples: 11850
  download_size: 442261896
  dataset_size: 677870272
- config_name: Joint
  features:
  - name: id
    dtype: large_string
  - name: task
    dtype: large_string
  - name: subtask
    dtype: large_string
  - name: q_type
    dtype: large_string
  - name: question
    dtype: large_string
  - name: answer
    dtype: large_string
  - name: referrable
    dtype: bool
  - name: visible
    dtype: bool
  - name: image
    dtype: image
  - name: image_alt
    dtype: image
  - name: cell
    dtype: large_string
  - name: target_layer
    dtype: large_string
  - name: img_center
    dtype: large_string
  - name: feature
    dtype: large_string
  splits:
  - name: train
    num_bytes: 495980580
    num_examples: 321
  - name: test
    num_bytes: 495980580
    num_examples: 321
  download_size: 988185534
  dataset_size: 991961160
- config_name: Spatial_Reasoning
  features:
  - name: id
    dtype: large_string
  - name: task
    dtype: large_string
  - name: subtask
    dtype: large_string
  - name: q_type
    dtype: large_string
  - name: question
    dtype: large_string
  - name: answer
    dtype: large_string
  - name: referrable
    dtype: bool
  - name: visible
    dtype: bool
  - name: image
    dtype: image
  - name: image_alt
    dtype: image
  - name: cell
    dtype: large_string
  - name: target_layer
    dtype: large_string
  - name: img_center
    dtype: large_string
  - name: feature
    dtype: large_string
  splits:
  - name: train
    num_bytes: 1764549611
    num_examples: 400
  - name: test
    num_bytes: 1764549611
    num_examples: 400
  download_size: 3518112486
  dataset_size: 3529099222
configs:
- config_name: Application
  data_files:
  - split: train
    path: Application/train-*
  - split: test
    path: Application/test-*
- config_name: ENC_Perception
  data_files:
  - split: train
    path: ENC_Perception/train-*
  - split: test
    path: ENC_Perception/test-*
  default: true
- config_name: Joint
  data_files:
  - split: train
    path: Joint/train-*
  - split: test
    path: Joint/test-*
- config_name: Spatial_Reasoning
  data_files:
  - split: train
    path: Spatial_Reasoning/train-*
  - split: test
    path: Spatial_Reasoning/test-*
language:
- en
---

# ENC-QA

ENC-QA evaluates vision-language model (VLM) understanding of Electronic Navigational Charts (ENCs). The benchmark measures whether models can interpret standardized nautical-chart symbols and text, reason about spatial relationships among chart features, combine ENC perception with spatial reasoning, and make navigation-relevant judgments from rendered ENC imagery.

ENC-QA is created from [National Oceanic and Atmospheric Administration (NOAA)](https://nauticalcharts.noaa.gov/charts/noaa-enc.html) ENCs encoded using the [IHO S-57 standard](https://iho.int/en/s-57-standards-in-force) and rendered using [OpenCPN](https://opencpn.org/). The benchmark contains approximately 56,000 questions derived from 6,835 NOAA ENCs.

The benchmark consists of four tasks in increasing complexity:

1. **ENC Perception:** recognize and interpret symbols, objects, text, and symbol-text associations visible in an ENC render.
2. **Spatial Reasoning:** reason about topology, distance, and direction between chart features.
3. **Joint Perception and Spatial Reasoning:** identify relevant chart features and use them jointly to answer spatial questions.
4. **Application:** synthesize perceptual and spatial information into navigation-relevant judgments.

## Data organization

The data directory follows the pattern:

```bash
ENC-QA/
├── data/
│   ├── ENC_Perception/
│   │   ├── isolated_symbol_recognition/
│   │   │   ├── questions.csv
│   │   │   └── images/
│   │   ├── incontext_symbol_recognition/
│   │   │   ├── questions.csv
│   │   │   └── images/
│   │   ├── symbol_text_association/
│   │   │   ├── questions.csv
│   │   │   └── images/
│   │   └── ...
│   ├── Spatial_Reasoning/
│   │   └── ...
│   ├── Joint/
│   │   └── ...
│   └── Application/
│       └── ...
└── README.md
```

The image path for each question is stored relative to the directory containing `questions.csv`. For example:

```text
images/US1EEZ1M_encperception_756619221_lat15.060136_lon145.793954_labeled.png
```

## Task hierarchy

ENC-QA uses three fields to identify where an example belongs in the benchmark:

- `task`: broad benchmark level.
- `subtask`: specific task within that level.
- `q_type`: fine-grained question or portrayal type.

For example, an in-context symbol-recognition example can be labeled as:

```text
task     = enc_perception
subtask  = incontext_symbol_recognition
q_type   = ADMARE/ADMARE
```

A spatial-reasoning example can be labeled as:

```text
task     = spatial_reasoning
subtask  = topological
q_type   = within
```

`q_type` should be treated as the fine-grained category produced by the question-generation pipeline. Its format depends on the task.

## Task definitions

### ENC Perception

ENC Perception evaluates whether a model can correctly identify and interpret individual chart elements.

The task includes:

- **Isolated Symbol Recognition:** identify an ENC object or symbol from an isolated or localized visual example.
- **In-Context Symbol Recognition:** identify an ENC object within a rendered chart scene.
- **Symbol-Text Association:** associate a chart symbol or feature with the corresponding visible text or label.

### Spatial Reasoning

Spatial Reasoning evaluates whether a model can reason about relationships between multiple chart features.

The task includes:

- **Topological**
  - within
  - touch
  - overlap
  - cross
  - equal
- **Distance**
  - quantitative distance
  - qualitative distance
- **Direction**
  - cardinal / 8-direction reasoning
  - exact bearing reasoning

To isolate spatial reasoning capability, the relevant objects or features are referred with neutral visual identifiers such as **red** and **blue**, rather than by their ENC object classes.

### Joint Perception and Spatial Reasoning

Joint tasks require both ENC visual grounding and spatial reasoning. A model must identify the relevant ENC features and then reason over their spatial relationship.

These tasks are designed to test whether a model can successfully combine the ENC-perception and spatial-reasoning capabilities that are evaluated separately in the lower-level tasks.

### Application

Application tasks evaluate whether a model can synthesize multiple ENC cues into a navigation-relevant judgment. These questions can require combining navigational aids, restricted areas, hazards, traffic routes, object attributes, and spatial relationships.

Application tasks include:

- **Travel Direction:** determine the intended or appropriate direction of travel from information visible on the chart.
- **Travel Permission:** determine whether travel or entry is permitted in the relevant chart area.
- **Change Criticality:** identify changes between two versions of a chart and determine whether those changes are critical to navigational safety.

Application examples may rely on both visible chart evidence and ENC feature attributes. The `visible` metadata field indicates whether the information needed to recover the answer is visually available in the rendered image.

## CSV schema

The current schema uses the following 15 columns:

| Column | Description |
|---|---|
| `id` | Unique string identifier for the example |
| `task` | Broad benchmark task, e.g. `enc_perception`. |
| `subtask` | Specific benchmark subtask, e.g. `incontext_symbol_recognition`. |
| `q_type` | Fine-grained question or portrayal category, e.g. `ADMARE/ADMARE` or `within`. |
| `question` | Natural-language prompt given to the model. |
| `answer` | Natural-language ground-truth answer. |
| `answer_alt` | Alternate ground-truth answer using the S-57 object code, portrayal code, abbreviation, or other task-specific coded representation when applicable. |
| `referrable` | Boolean indicating whether the target can be referred by a name or object type. |
| `visible` | Boolean indicating whether the information required for the answer is considered visible in the rendered image. |
| `image` | Relative path to the primary image used for the question. |
| `image_alt` | Relative path to a paired alternate rendering when available. In most cases, this is the corresponding scene without the target highlight/annotation. |
| `cell` | NOAA ENC cell identifier, e.g. `US1EEZ1M`. |
| `target_layer` | S-57/ENC layer containing the target source feature, e.g. `ADMARE`, `COALNE`, or `LNDELV`. |
| `img_center` | Image-center coordinate in WGS84 stored as `[longitude, latitude]`. |
| `feature` | JSON-encoded list containing structured metadata for the source feature or features used to generate the question. |

Not every task necessarily requires every optional field. Fields such as `answer_alt`, `target_layer`, `img_center`, or `feature` may be empty when they are not applicable to a particular task.

### Example row

A simplified in-context symbol-recognition example is:

```text
id:
US1EEZ1M_encperception_756619221_lat15.060136_lon145.793954

task:
enc_perception

subtask:
incontext_symbol_recognition

q_type:
ADMARE/ADMARE

question:
The image displays a symbol used to represent ENC data on an ECDIS.
What type of object or feature does the highlighted area portray?
Give the final answer in <answer> </answer> tags.

answer:
Administration Area (Named)

answer_alt:
ADMARE

visible:
True

image:
images/US1EEZ1M_encperception_756619221_lat15.060136_lon145.793954_labeled.png

image_alt:
images/US1EEZ1M_encperception_756619221_lat15.060136_lon145.793954.png

cell:
US1EEZ1M

target_layer:
ADMARE

img_center:
[145.79395375541344, 15.06013575]
```

## Feature metadata

The `feature` column stores a JSON-encoded list containing metadata for the source feature or features associated with an example.

Common fields include:

| Field | Description |
|---|---|
| `layer` | ENC layer associated with the feature. |
| `geometry_type` | Geometry class, such as `point`, `line`, or `polygon`. |
| `geometry` | Serialized geometry representation for the source feature. |
| `fid` | Feature identifier used by the source data-processing pipeline. |
| `rcid` | S-57 record identifier. |
| `object_name` | Human-readable object-class name when available. |
| `target_label` | Target label produced for the task. |
| `target_label_source` | Generator metadata describing how `target_label` was selected. |
| `portrayal_name` | Portrayal or symbol identifier associated with the target. |
| `attributes` | Serialized dictionary containing source S-57 attributes. |

For example:

```json
[
  {
    "layer": "ADMARE",
    "geometry_type": "polygon",
    "fid": 141,
    "rcid": 142,
    "object_name": "Administration Area (Named)",
    "target_label": "ADMARE",
    "target_label_source": "object_class_for_polygon_task",
    "portrayal_name": "ADMARE",
    "geometry": "...",
    "attributes": "..."
  }
]
```

`geometry` and `attributes` may themselves be stored as serialized strings rather than recursively expanded JSON objects.

## Primary and alternate images

Some ENC-QA tasks contain two distinct images for the image fields:

- `image`: the primary image associated with the question.
- `image_alt`: a paired alternate rendering of the same scene.

In most cases, the primary image contains the target highlight or annotation used by the question. The alternate image preserves the corresponding chart scene without that target annotation. It may still contain normal chart labels and other OpenCPN-rendered content.

## Dataset statistics

| Configuration | Total |
|---|---:|
| **ENC Perception** | TODO |
| &emsp;Isolated Symbol Recognition | TODO |
| &emsp;In-Context Symbol Recognition | TODO |
| &emsp;Symbol-Text Association | TODO |
| **Spatial Reasoning** | TODO |
| &emsp;Topological | TODO |
| &emsp;Distance | TODO |
| &emsp;Direction | TODO |
| **Joint Perception and Spatial Reasoning** | TODO |
| **Application** | TODO |
| &emsp;Travel Direction | TODO |
| &emsp;Travel Permission | TODO |
| &emsp;Change Criticality | TODO |
| **Total** | **~56k** |

## Data sources

### NOAA Electronic Navigational Charts

ENC-QA is derived from NOAA Electronic Navigational Charts. ENCs are vector nautical charts encoded using the IHO S-57 data model and contain navigational objects, attributes, geometry, and chart metadata.

**Data checkpoint:** June 1, 2026.

### Chart rendering

ENC source data are converted into rendered chart imagery for the visual benchmark using OpenCPN. Structured ENC information is used to identify source features, attributes, and geometric relationships, while the rendered chart is provided as the visual model input.

## Example usage

### Reading the CSV

```python
from pathlib import Path

import pandas as pd

task_root = Path(
    "/path/to/ENC-QA/data/ENC_Perception/incontext_symbol_recognition"
)

questions = pd.read_csv(task_root / "questions.csv")

row = questions.iloc[0]

image_path = task_root / row["image"]
image_alt_path = (
    task_root / row["image_alt"]
    if pd.notna(row.get("image_alt"))
    else None
)

print(row["id"])
print(row["task"])
print(row["subtask"])
print(row["q_type"])
print(row["question"])
print(row["answer"])
print(row.get("answer_alt"))
print(image_path)
print(image_alt_path)
```