---
license: mit
task_categories:
  - image-feature-extraction
  - robotics
tags:
  - scene-graph-generation
  - ai2thor
  - embodied-ai
  - adversarial-attacks
size_categories:
  - 1K<n<10K
---

# NSAP sim data

Ground-truth scene graphs rendered from AI2-THOR, built for the NSAP paper on adversarial
attacks against the symbolic grounding of scene-graph models.

Every frame ships an instance segmentation map and a scene graph with per-object
categories, attributes and pairwise relations, all read from the simulator rather than
predicted. That is the point of the set: it lets you score a scene-graph model against
exact ground truth, and it lets an attack be trained against precise object masks with no
detector in the loop.

Code: https://github.com/MasoudJTehrani/NSAP

## Getting it

```python
from huggingface_hub import snapshot_download
snapshot_download("MasoudJTehrani/NSAP-sim-data", repo_type="dataset", local_dir="dataset/sim")
```

Or, from the code repository, which also lays out `kitchen_ebnav/` where the tooling
expects it:

```bash
python NSAP.py --with-dataset
```

Pin a revision if you care about exact reproducibility. The NSAP code pins one by default.

## Layout

```
manifest.json                       202 scenarios: id, scene, target category, frame count, source
scenarios/<scenario_id>/
    meta.json                       scenario id, FloorPlan, seed, target object, frame count
    scene_graph.json                the ground truth, per frame (see below)
    frames/NNN.png                  RGB, 600x600
    instance/NNN.png                instance segmentation, colour-keyed to object ids
kitchen_ebnav/                      12 extra scenarios, scene-matched to EB-Navigation episodes
```

`scene_graph.json` gives, for each frame, every visible object with its integer id, category,
AI2-THOR object id, instance colour, bounding box and attribute list, plus the frame's
relation edges.

## What is in it

202 scenarios, 2389 frames, 30 FloorPlans, 75 distinct object categories. Roughly 12
viewpoints per scenario, orbiting a designated target object.

| subset | scenarios | built for |
|---|---|---|
| `pilot_*` | 161 | general coverage: apple, bowl, coffee machine, fridge, kettle, microwave, pot, toaster |
| `tv_*` | 30 | the attribute head, televisions in `displaying` / `showing` states |
| `pot_*` | 11 | the relation head, pots staged in controlled `on` configurations |

Attribute vocabulary: `open`, `closed`, `clean`, `off`, `on`, `displaying`, `showing`, `filled`.
Relation vocabulary: `near`, `on`, `in`.

`kitchen_ebnav/` is separate on purpose. Its 12 scenarios are framed on the exact object
instances that EB-Navigation episodes target, so a patch trained there can be injected into
the live benchmark. Targets are pot (6), toaster (3), kettle (2) and coffee machine (1).

## Generation

Rendered with [SceneGraphBench](https://github.com/MasoudJTehrani/SGDS_AI2_THOR) at seed 0,
600x600, 12 viewpoints per target, minimum visible area 600 px, with object states
randomised and materials held fixed. `manifest.json` carries the exact generator config.

## Citation

If you use this, please cite the NSAP paper. AI2-THOR should be cited separately.