The Assembled Panel Dataset (APD) is released for non-commercial research and educational use only, under the APD Research-Only License (see LICENSE). By requesting access you agree to these terms, not to redistribute the images or annotations, and to cite the paper below in any work that uses the dataset.

Assembled Panel Dataset (APD)

The Assembled Panel Dataset (APD) is a real industrial inspection dataset for anomaly detection and localization. Each image shows the same assembled electrical panel, with a PCB enclosure, power and EV-port cabling, a terminal block, and a current sensor, captured under a fixed setup. Anomalous panels contain one or more of 14 assembly errors, such as swapped or missing cables, missing screws, labels, caps, or rubber rings, incorrect wire routing, and PCB damage. Every error is annotated at pixel level.

The APD is the dataset of Scenario 3 (industrial robustness) of the TGAD benchmark, introduced in A Structured Benchmark for Text-Guided Anomaly Detection: When Language Stops Conditioning the Decision (BMVC 2026). There it is evaluated in two settings:

Standard AD: every anomalous image is positive, and its ground truth is the union of all of its defect masks;
Per-Error AD: for each of the 14 categories, its positive images are compared with all normal images, using only that category's mask.

At a glance


Images	142 RGB PNG, 1920 x 1080
Normal (defect-free)	89
Anomalous	53 (10 of them contain more than one category)
Anomaly categories	14 (`err_1` ... `err_14`)
Annotations	75 polygon instances, 64 image-category assignments, COCO format
Splits	test only; there is no training split

Structure

normal/                                     89 defect-free images
anomalous/                                  53 anomalous images
assembled_panel_annotations.json            COCO annotations of the anomalous images
ImageAcquisitionProtocolAssembledPanel.pdf  Image acquisition protocol
error_code_for_anomalies.pdf                Representative images of each error code
LICENSE                                     APD Research-Only License

Normal images have no annotations; their ground-truth mask is empty.

ImageAcquisitionProtocolAssembledPanel.pdf describes the camera, the enclosure, and the five tested acquisition setups; the released images were captured with Setup 5. error_code_for_anomalies.pdf shows representative images of each of the 14 error codes.

Anomaly categories

Category A_k is the COCO category named err_k. The COCO category ids are not in this order (for example, id 1 is err_10), so match categories by name.

ID	Code	Category	Visual signature	Images
A1	`err_1`	swapped power-cable colors	color permutation at left/right power ports	5
A2	`err_2`	swapped large cables in EV ports	green/blue large cables exchanged between left/right EV ports	4
A3	`err_3`	swapped small cables in EV ports	orange/white small cables exchanged between left/right EV ports	6
A4	`err_4`	swapped same-color IO power cables	upper/lower same-color power cables exchanged at terminal block	1
A5	`err_5`	misplaced label	identification label present but in non-standard position	5
A6	`err_6`	missing label	absent identification sticker on PCB black box	5
A7	`err_7`	missing rubber ring near mounting screw	absent ring around panel-mounting screw head	5
A8	`err_8`	missing red cap	absent red protective cap near EV port	7
A9	`err_9`	missing power-connection wire(s)	absent wire at top/bottom power port	5
A10	`err_10`	missing wire(s) on EV port connections	absent wire at left/right EV port connection	7
A11	`err_11`	missing screw(s)	absent screw at panel or PCB mounting position	5
A12	`err_12`	incorrect wire routing through current sensor	wire count through right current sensor differs from two	6
A13	`err_13`	scratched or broken PCB	surface scratch, fracture, or missing component on PCB	1
A14	`err_14`	unplugged wire(s) on current sensor	disconnected wire at the right current sensor	2

"Images" counts the anomalous images that contain the category. An image with several categories is counted once for each.

Annotation format

assembled_panel_annotations.json is a COCO detection/segmentation file exported from Label Studio:

images: 53 entries with width, height, and file_name. The file_name values are absolute Label Studio paths, so match them to the image files by basename.
annotations: polygon segmentation, bbox as [x, y, width, height], area, category_id, image_id.
categories: err_1 ... err_14.

Image file names also contain notes written at capture time. They are not ground truth; use the COCO file.

Usage

Access is gated: log in to Hugging Face and accept the terms on this page first. Then:

hf auth login   # or set HF_TOKEN
hf download StefanoSamele-PoliMi/assembled-panel-dataset --repo-type dataset --local-dir apd

Or in Python:

from huggingface_hub import snapshot_download

root = snapshot_download("StefanoSamele-PoliMi/assembled-panel-dataset", repo_type="dataset", local_dir="apd")

To rasterize the masks of one image (requires pycocotools):

import os
from pycocotools.coco import COCO

coco = COCO("apd/assembled_panel_annotations.json")
by_name = {os.path.basename(img["file_name"]): img["id"] for img in coco.dataset["images"]}
image_id = by_name["sample_right_12_err12_2025-09-30_15-04-31.png"]
masks = {coco.cats[a["category_id"]]["name"]: coco.annToMask(a) for a in coco.loadAnns(coco.getAnnIds(imgIds=image_id))}

masks maps each category name to its binary mask. If an image has two annotations of the same category, keep the union of their masks.

License

The APD is released under the APD Research-Only License (see LICENSE). It permits non-commercial research and educational use only. Redistribution is not allowed, and publications that use the dataset must cite the paper below. For any other use, contact the authors.

Citation

@inproceedings{samele2026tgad,
  title     = {A Structured Benchmark for Text-Guided Anomaly Detection: When Language Stops Conditioning the Decision},
  author    = {Samele, Stefano and Lomurno, Eugenio and Jovanovic, Teodora and Shivakumar Manohar, Sanjay and Crivellaro, Alberto and Matteucci, Matteo},
  booktitle = {British Machine Vision Conference (BMVC)},
  year      = {2026}
}

Downloads last month: 187

Total file size:

354 MB