README.md

---
tags:
- smart-manufacturing
- sft
- industrial
- vision
license: other
pretty_name: "189"
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

Industrial 3D object recognition & 6D pose (multi-modal gray+depth+masks; per-instance; DATA-ONLY). Category **B**, task **T-B5b**, 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) |
| `annot` | null (data-only) | the answer — for this dataset: **none** — this is a DATA-ONLY release: `query` is empty and `annot` is `null` (no task framing yet). Each row is one object instance; its four images (scene gray, scene depth, instance mask, instance visible mask) are in the `images` column, and the full ground truth (6D pose `cam_R_m2c`+`cam_t_m2c`, 2D `bbox_obj`/`bbox_visib`, visibility, `cam_K`, model geometry) is in `metadata`. See **Modalities, GT & possible tasks** below for what this data supports (incl. our suggested 2D-detection framing) |
| `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] | *(D21 only)* multi-region masks |

## Modalities, ground truth & possible tasks

**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.

**This release is DATA-ONLY (no query, no annot).** `query` is an empty string and `annot` is `null` on
purpose — we ship the raw multi-modal data with rich ground truth so a task can be framed later. **Only the
`validation` split is included**: it is the source's **val** scenes, the only ones with released ground
truth (**54 images / 123 object instances / all 28 objects**). The source's 721-image **test** set has its
GT withheld on the evaluation server and is **not** included.

**One row = one object instance, with all its images.** Each record is a single annotated instance and its
`images` column holds **four images 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** — the instance's full object mask (0/255);
4. **mask_visib** — the instance's *visible* mask (0/255).
The scene gray/depth are shared by every instance in that image (duplicate bytes are deduplicated in the
parquet). `image` (scalar) = the gray image.

**All ground truth is in `metadata`** (JSON): `obj_id` (1–28), **6D pose** `cam_R_m2c` (3×3 row-major
rotation) + `cam_t_m2c` (translation, mm), 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). A text VLM cannot emit a rotation matrix, so the 6D pose is kept as data/GT, not framed as a
task.

**Tasks this data supports.**
- *Original benchmark task (Drost et al. 2017):* **6-DoF pose estimation** and **3D object detection** from
  gray + depth (evaluated with pose-error metrics such as ADD/VSD). The pose + camera + model GT here are
  exactly what such a task needs.
- *Instance segmentation / object presence & counting:* the per-instance masks + `obj_id` list support
  segmentation and "which/how many of the 28 parts are in this bin".
- **2D object detection — suggested by us (this project), NOT part of the original benchmark.** Because the
  6D pose is not text-friendly, a natural VLM framing is to detect each instance's 2D box: input the gray
  (+ optionally depth), output one line per instance `obj_<id>,[x, y, w, h]` using `bbox_visib`. Here
  `obj_<id>` is a *specific CAD-modelled part* (visually consistent, so learnable), just numbered rather
  than named. This framing is our suggestion for downstream SFT use; it is not an official ITODD task and
  is not baked into this release (which stays data-only).

**Note on the object ids.** The 28 objects are identified by number (`obj_1`..`obj_28`); ITODD does not give
them semantic names. Unlike an arbitrary code, each id is one fixed physical part (with a CAD model), so it
is a well-defined fine-grained class.

## 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 script: `189/convert_d89.py`, published with `publish/push_to_hf.py`, both 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.