README.md

---
pretty_name: OOD-PerceptionBench asset pack
license: other
license_name: mixed-cc-by-4.0-and-cc-by-nc-4.0
license_link: https://github.com/farrosalferro/OODPerceptionBench/blob/master/assets/ATTRIBUTION.md
tags:
  - carla
  - autonomous-driving
  - out-of-distribution
  - benchmark
  - 3d-assets
  - simulation
---

# OOD-PerceptionBench — asset pack v0.9

Cooked CARLA 0.9.15 assets for **OOD-PerceptionBench**, a closed-loop driving benchmark
that separates *visual* (appearance) out-of-distribution shift from *geometric*
(shape/size) shift.

**Code, routes, baseline records and documentation:**
<https://github.com/farrosalferro/OODPerceptionBench> (branch `master`).
This repository holds only the binary assets, which are too large for GitHub.

> **Version stamp:** asset pack **v0.9**, corresponding to **arXiv v1** of the paper.
> v1.0 will add replacement assets for the props that cannot be redistributed today, and
> the numbers it produces will *not* be interchangeable with v0.9 numbers. Always record
> which pack version you used.

---

## ⚠ Read this before you use it

**This pack contains six of the benchmark's eighteen OOD props.** The other twelve are not
redistributable (marketplace licences that forbid standalone redistribution and AI use;
two free downloads whose uploaders did not own the underlying IP). They are specified
**dimensionally** instead, and the repository ships classifier notebooks so you can check
your own substitutes against the same admissibility rule the paper uses.

**With this pack you can run 237 of the 475 routes.** See §4.

**One asset is NonCommercial.** `walker.pedestrian.firefighter` is **CC BY-NC 4.0**. It is
packaged separately so you can leave it out. See §3.

**A missing CARLA asset fails silently.** If a blueprint ID is not registered, the
benchmark harness substitutes a different actor and the route finishes with a plausible
Driving Score. Run `tools/verify_pack.py` after installing. This is not optional advice.

---

## 1. Contents

| Blueprint ID | Category | Shift level | Author | Licence |
|---|---|---|---|---|
| `walker.pedestrian.astronaut` | pedestrian | visual | Antropik | CC BY 4.0 |
| `walker.pedestrian.firefighter` | pedestrian | visual | KIFIR | ⚠ **CC BY-NC 4.0** |
| `walker.pedestrian.deliveryrobot` | pedestrian | geometric | Bento (@gostbento) | CC BY 4.0 |
| `walker.pedestrian.boar` | pedestrian | geometric | AnimalMesh 3D | CC BY 4.0 |
| `static.prop.concreteroadbarrier` | static | geometric | widthRider | CC BY 4.0 |
| `static.prop.roadclosedbarricade` | static | geometric | exiS7-Gs | CC BY 4.0 |

Plus one modified CARLA base-content asset, `WalkerFactory` (CC BY, © CARLA Simulator
authors) — the four walkers cannot register without it. Full licence text and links:
`ATTRIBUTION.md`. Why the factory has to ship, and what it costs: `WALKERFACTORY_DECISION.md`.

```
ood-perceptionbench-props-v0.9.tar.gz          15.8 MB   CC BY 4.0
ood-perceptionbench-walkers-ccby-v0.9.tar.gz   38.4 MB   CC BY 4.0   (includes WalkerFactory)
ood-perceptionbench-walkers-ccbync-v0.9.tar.gz 112.3 MB  CC BY-NC 4.0
                                              --------
                                               166.6 MB download · 185.5 MB installed · 195 files
```

---

## 2. Install

Requires an official **CARLA 0.9.15** Linux build. Full instructions, including uninstall
and a `tar` pitfall that can silently skip the factory, are in `INSTALL.md`.

```bash
cd /path/to/CARLA_0.9.15
sha256sum -c SHA256SUMS
mkdir -p Import && cp ood-perceptionbench-*-v0.9.tar.gz Import/
./ImportAssets.sh

# then, with a CARLA server running:
python3 tools/verify_pack.py --port 2000
```

`ImportAssets.sh` is CARLA's own installer; the tarballs are rooted at `CarlaUE4/` so they
extract straight into the build.

**The pack overwrites `CarlaUE4/Content/Carla/Blueprints/Walkers/WalkerFactory.{uasset,uexp}`.**
Install into a copy of CARLA dedicated to this benchmark, and only over 0.9.15.

---

## 3. The NonCommercial asset

`walker.pedestrian.firefighter`, by **KIFIR**, is **CC BY-NC 4.0**: no commercial use of
the model, of derivatives, of rendered frames, or of datasets built from them.

It lives in its own tarball. If your use is commercial, **do not install
`ood-perceptionbench-walkers-ccbync-v0.9.tar.gz`** and run the verifier with
`--without-nc`. Everything else in the pack is CC BY 4.0. The cost of omitting it is 18
pedestrian routes.

There is no IP defect here — it is the author's original work — and it is deliberately not
being replaced.

---

## 4. Route coverage

| Category | Runnable with this pack | Total | Gap |
|---|---|---|---|
| static | **30** | 70 | 40 routes need 4 non-redistributable props |
| pedestrian | **126** | 162 | 36 routes need `soldier` / `wheelchair` |
| vehicle | **81** | 243 | 162 routes need 6 non-redistributable vehicles |
| **total** | **237** | **475** | 238 |

145 of the 237 are the benchmark's `base`-level routes, which use native CARLA blueprints
and need no pack at all. This pack unlocks the other 92 shift-level routes: 18 each for
astronaut, firefighter, boar and deliveryrobot, 10 each for concreteroadbarrier and
roadclosedbarricade.

The 238 remaining routes are **not runnable at v0.9**. Running them anyway yields
substituted-actor results that look normal and mean nothing.

Per-route baseline records for the full 475 across 17 models are published in the code
repository — verifying the paper's claims needs no GPU and no assets.

---

## 5. Phantom blueprint IDs

After installing, the blueprint library will advertise nine walker IDs whose content is not
in this pack: `soldier`, `wheelchair`, `ball`, `caneman`, `cow`, `crutcheswoman`, `deer`,
`labrador`, `tire`. They come from the shipped `WalkerFactory`, which was cooked from a
build containing more assets than are redistributable.

All nine are **unspawnable** — `try_spawn_actor` returns `None` and `spawn_actor` raises.
They cannot silently give you the wrong prop, but they are not usable. `verify_pack.py`
asserts that none of them spawns. Details and the reasoning: `WALKERFACTORY_DECISION.md`.

---

## 6. Files in this repository

```
README.md                     this file
INSTALL.md                    install, verify, uninstall, rebuild
ATTRIBUTION.md                per-asset authors, licences, source links
WALKERFACTORY_DECISION.md     why WalkerFactory ships and what it costs
SHA256SUMS                    checksums of the three tarballs
MANIFEST.tsv                  every shipped file: path, sha256, size, asset, licence
dist/                         the three tarballs
tools/verify_pack.py          post-install verification (run it)
tools/goldens.json            reference bounding boxes for the six assets
build/build_asset_pack.sh     reproduces this pack from a CARLA build
build/make_manifest.py        emits SHA256SUMS and MANIFEST.tsv
build/EXCLUSIONS.tsv          files present in the source build and deliberately not shipped
```

---

## 7. Citation

Cite the OOD-PerceptionBench paper (see `CITATION.cff` in the code repository), credit
CARLA for the simulator content, and reproduce the per-asset attribution in
`ATTRIBUTION.md` — CC BY 4.0 requires it.