---
license: cc-by-4.0
task_categories:
  - object-detection
tags:
  - robotics
  - manipulation
  - simulation
  - referring-expressions
size_categories:
  - 1K<n<10K
---

# LIBERO for detection

A small multi-class object detection set rendered from **LIBERO**, for an
undergraduate robot perception module at the National University of Singapore.

Nothing here is a new benchmark. LIBERO is a lifelong-robot-learning benchmark
by Liu, Zhu, Zhu, Stone and colleagues; the scenes, objects, task descriptions
and initial states are theirs, at <https://libero-project.github.io>, and that
is the thing to cite.

## What it is

LIBERO's scenes are built from description files and rendered, so the simulator
also produces an instance segmentation in which every mask carries the name of
the model that made it — `akita_black_bowl_2` is a labelled instance of
`akita_black_bowl`. Thirty-one classes come out of that, and nobody drew a box.

Each frame also carries its task's language instruction and the instance that
instruction refers to, which LIBERO states in the task file as
`(:obj_of_interest ...)`. Some scenes hold three identical bowls and the
instruction names one of them, so the referring expression is labelled rather
than guessed.

## How it was rendered

The scenes are LIBERO's. Four decisions in the rendering are worth stating:

- **The scene settles before the shutter.** Objects are dropped a few
  centimetres and gravity is allowed to finish, about six simulator steps.
  Rendering at once boxes a bowl where it is not going to be.
- **The layouts are LIBERO's own fifty validated initial states per task**,
  not re-sampled placements — its sampler can leave an object falling off a
  shelf.
- **The camera moves slightly in every frame**, by up to 8 cm, 4° and 3° of
  field of view, around LIBERO's one fixed `agentview` pose.
- **Half the frames have the arm over the table.** It is driven there between
  the two renders of a layout, and the frame is dropped if anything it passed
  over shifted more than 2 cm. Occlusion is wanted; rearranging the scene is
  not.

Rendered at 512×512 and resampled to 256×256.

## The split

**By scene.** Six of the twenty-two rooms are held out whole: `KITCHEN_SCENE1`,
`KITCHEN_SCENE4`, `KITCHEN_SCENE6`, `KITCHEN_SCENE9`, `LIVING_ROOM_SCENE3`,
`STUDY_SCENE2`. An earlier version split by layout within a task instead, which
leaves near-duplicates on both sides: LIBERO's initial states move objects by
only a few centimetres.

Twenty-two of the thirty-one classes appear in the held-out rooms; the other
nine are trained on and not scored. A split in which all thirty-one appear on
both sides does not exist on this object set — several classes live in only
four or five tasks.

So a number from this set measures **unseen rooms and unseen arrangements, with
objects the detector has met elsewhere**. It does not measure transfer to new
object categories, and it does not measure transfer to a new viewpoint: every
frame on both sides is that same camera, with the jitter above.

## Contents

```
images/train/<task>__<layout>_<pose>.jpg     2154 frames, 256x256
images/test/ …                                718 frames
train.json / test.json
```

Each record:

| field | |
|---|---|
| `file` | the image |
| `suite`, `task` | which LIBERO task it came from |
| `language` | the instruction that task was recorded for |
| `target` | the instance(s) the instruction is about |
| `objects[]` | `instance`, `cls`, `label` (1–31), `box` as x1 y1 x2 y2, `area` |

| | train | test |
|---|---|---|
| frames | 2154 | 718 |
| boxes | 10935 | 2820 |
| scenes | 16 | 6 |
| classes | 31 | 22 in the held-out rooms |
| objects per frame, median | 5 | 4 |
| box side, median | 39 px | 43 px |
| frames whose referred object is visible | 98% | 99% |
| download | 34 MB | |

## Licence

CC BY 4.0, following LIBERO. If you use it, cite LIBERO.