README.md

---
tags:
- smart-manufacturing
- sft
- industrial
- vision
license: other
pretty_name: "189-agent"
extra_gated_fields:
  Name: text
  Affiliation: text
  Intended use: text
extra_gated_prompt: >-
  This dataset is released for **research use**. Access is reviewed and granted
  **manually** by the maintainers. Please state your name, affiliation, and intended use.
---

# 189-agent

Industrial 6-DoF object pose estimation (per-instance; gray+depth+mask input, CAD .ply provided; graded with ADD/VSD). Reserved for an AGENTIC setting, not VLM SFT. Category **B**, task **T-B4**, in the unified Smart-Manufacturing SFT schema.

> The repository name is an internal task code. See **Provenance** below for the underlying dataset.

## Records

**123** records (validation=123). 

## Unified SFT schema

| field | type | meaning |
|---|---|---|
| `query` | str | the question / instruction (model input) |
| `image` | Image | the input image (bytes embedded); for multi-image rows, a preview of the first view |
| `images` | list[Image] | *(multi-image rows)* all input views / modalities for the row, bytes embedded |
| `annot` | str (JSON pose) | the answer — for this dataset: a JSON object `{"cam_R_m2c": [9 numbers = row-major 3x3 rotation], "cam_t_m2c": [3 numbers = translation in mm]}` — the target instance's 6-DoF pose ground truth, a CONTINUOUS geometric quantity graded with ADD/VSD pose-error metrics (NOT string match). `images` = [scene gray, scene depth (uint16 mm), the instance's visible mask]; the object's CAD model is `metadata.model_ply` (uploaded under `assets/models/`). See **Task, inputs & split** below |
| `reasoning` | null | no native CoT in these datasets |
| `cate` | "B" | SFT category |
| `task` | "T-xx" | unified task id |
| `metadata` | str (JSON) | split, provenance, `image_path`, `image_sha256` (dedup key) |
| `mask` | Image \| null | *(T-B1/T-B2 only)* the pixel ground-truth mask, bytes embedded |
| `masks` | list[Image] | *(multi-image T-B1 / D21)* per-view masks aligned with `images` (None where a view has no defect), or multi-region masks |

## Task, inputs & split

**What this is.** MVTec ITODD (Drost, Ulrich, Bergmann, Härtinger, Steger — *Introducing MVTec ITODD: A
Dataset for 3D Object Recognition in Industry*, ICCV Workshops 2017) — a benchmark for **industrial 3D
object recognition and 6-DoF pose estimation**. 28 rigid industrial parts (each with a CAD model), imaged
in bins with a **grayscale + depth** industrial 3D sensor, in **BOP** format.

**Task (this release): 6-DoF pose estimation.** One row = one object instance. The `query` (our own
template) identifies the target — object `obj_<id>`, its 2D bounding box `[x, y, w, h]`, and its CAD model
(`metadata.model_ply`) — and asks for that instance's pose. `annot` is a JSON object
`{"cam_R_m2c": [9 numbers], "cam_t_m2c": [3 numbers]}`: the **3×3 row-major rotation R** and the
**translation t (mm)** of the object in the camera frame. This is a **continuous geometric** answer —
grade it with the standard **ADD / VSD** pose-error metrics, **not** string match. The instance is
*specified* (its bbox + visible mask are given), so localisation is an input, not part of the metric.

**Inputs — `images`, three in this fixed order** (`metadata.modalities`):
1. **gray** — the scene grayscale image (8-bit, 1280×960);
2. **depth** — the scene depth map, re-encoded to a **16-bit PNG in millimetres** (multiply by
   `metadata.depth_scale`; the original float `.tif` path is in `metadata.depth_raw_tif`);
3. **mask_visib** — the target instance's *visible* mask (0/255), locating the specified object.
The scene gray/depth are shared by every instance in that image (duplicate bytes are deduplicated in the
parquet). `image` (scalar) = the gray image. The object's **CAD model** ships as a `.ply` under
`assets/models/obj_<id:06d>.ply`, referenced per row by `metadata.model_ply` — it is what ADD/VSD needs.

**Other ground truth / metadata** (JSON): `obj_id` (1–28), 2D boxes `bbox_obj` / `bbox_visib`
([x, y, w, h]), `visib_fract`, pixel counts, camera intrinsics `cam_K` + `depth_scale`, and the object's
`model_geometry` (diameter, bbox dims in mm). `obj_<id>` is a *specific CAD-modelled part* (visually
consistent), just numbered rather than named — a well-defined fine-grained class.

**Split.** **Only `validation` is included** — the source's **val** scenes, the only ones with released
GT (**54 images / 123 object instances / all 28 objects**). The 721-image **test** set has its GT withheld
on the evaluation server and is **not** included.

**Intended use — agentic, not SFT.** A 6-DoF pose (a continuous 3x3 rotation + translation) is ill-suited
to plain text-output VLM supervised fine-tuning, so this dataset (`189-agent`) is **reserved for an agentic
setting** — e.g. a tool-using agent that calls a pose solver / renderer and is scored by ADD/VSD against
this ground truth — rather than direct model training.

## Provenance

Underlying dataset: **MVTec-ITODD**. Upstream license: **other (research use; MVTec ITODD, Drost et al., ICCV Workshops 2017)** (this card is `license: other`; respect the upstream terms). Converted read-only from the raw source into the unified schema; conversion code under `189/` (with `publish/push_to_hf.py`) in [`AI4Manufacturing/forge_model`](https://github.com/AI4Manufacturing/forge_model).

## Overlap / de-duplication (§8)

Only the val split (54 images / 123 instances) ships GT; the 721-image test set has GT withheld upstream and is excluded. Each record carries `metadata.image_sha256` so overlapping images can be kept entirely on one side of a train/eval split.