martinsawojide's picture
Add files using upload-large-folder tool
ac62cf0 verified
|
Raw History Blame Contribute Delete
7.32 kB
---
license: cc-by-4.0
pretty_name: webcheck synthetic print-defect test frames
task_categories:
- object-detection
annotations_creators:
- machine-generated
source_datasets:
- original
size_categories:
- 1K<n<10K
tags:
- synthetic
- anomaly-detection
- industrial-inspection
- blenderproc
- flexible-packaging
configs:
- config_name: A
data_files:
- split: test
path:
- "A/*.png"
- "A/metadata.jsonl"
- config_name: B
data_files:
- split: test
path:
- "B/*.png"
- "B/metadata.jsonl"
dataset_info:
- config_name: A
features:
- name: image
dtype: image
- name: camera
dtype: string
- name: has_defect
dtype: bool
- name: objects
struct:
- name: bbox
list:
list: float32
length: 4
- name: categories
list:
class_label:
names:
'0': streaks
'1': dot_skip
'2': haze
'3': mottling
'4': tunneling
'5': bubbles
'6': silvering
'7': wrinkles
- config_name: B
features:
- name: image
dtype: image
- name: camera
dtype: string
- name: has_defect
dtype: bool
- name: objects
struct:
- name: bbox
list:
list: float32
length: 4
- name: categories
list:
class_label:
names:
'0': streaks
'1': dot_skip
'2': haze
'3': mottling
'4': tunneling
'5': bubbles
'6': silvering
'7': wrinkles
---
# webcheck synthetic print-defect test frames
Synthetic camera frames of a printed flexible-packaging web with labelled print defects. These are the held-out test
frames that the webcheck demo plays on an NVIDIA Jetson Orin Nano: a FOMO-AD anomaly gate followed by a
YOLO-Pro detector, both trained in Edge Impulse. No frame here was used to train either model or to set any threshold.
## Dataset details
### Description
- **Curated by:** Martins Awojide
- **License:** [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)
- **Repository and write-up:** webcheck (links to follow)
Every frame is a render of one fictional product (all brands, logos and codes are placeholders) on a moving web, seen
by one of two cameras placed across the web (`left`, `right`). There are two placements, each its own configuration:
| Config | Placement | Film | Frames | Frames with a defect | Boxes | Classes present |
|---|---|---|---|---|---|---|
| `A` | press side, before lamination | outer PET with reverse-printed ink | 553 | 157 | 175 | 4 |
| `B` | slitter side, cured laminate | PET / reverse-printed ink / AlOx-coated PET / LLDPE, transflected light | 553 | 294 | 342 | 8 |
Counts are from the COCO files (counted 2026-10-31).
Boxes per class:
| Class | `A` | `B` |
|---|---|---|
| streaks | 48 | 48 |
| dot_skip | 34 | 34 |
| haze | 48 | 42 |
| mottling | 45 | 46 |
| tunneling | – | 48 |
| bubbles | – | 37 |
| silvering | – | 45 |
| wrinkles | – | 42 |
The demo's detector covers six classes (streaks, dot skip, mottling, tunneling, bubbles, silvering). Haze and wrinkles
are labelled here but the demo models do not detect them.
## Uses
- **Intended:** running the webcheck demo (`scripts/run_demo.sh`), and viewing or evaluating detectors of print
defects on synthetic frames.
- **Not intended:** training. This is a test split only; the train and validation splits are not published. It is not
a substitute for frames from a real press or slitter.
## Dataset structure
```
A/ 553 PNG metadata.jsonl manifest.jsonl _annotations.coco.json SHA256SUMS
B/ 553 PNG metadata.jsonl manifest.jsonl _annotations.coco.json SHA256SUMS
```
- **PNG frames:** 1920x1080 RGB, 0.5 mm per pixel. File names repeat between `A/` and `B/`; the folder is the
placement.
- **`metadata.jsonl`:** one line per frame, read by the `datasets` library: `file_name`, `camera`, `has_defect` and
`objects` (`bbox` as COCO `[x, y, width, height]` in pixels, `categories` as 0-based class indices).
- **`_annotations.coco.json`:** the same boxes in COCO form, with extra per-box fields (`visible_fraction`,
`truncated`, `area_mm2`, `web_extent_mm`, `attributes.severity_mm`). COCO category ids are 1-based and shared by
both placements: `A` uses ids 1–4, `B` uses 1–8.
- **`manifest.jsonl`:** one line per frame from the renderer (roll, frame index, camera, offsets on the web, seed).
The demo reads this file. Its `image_sha256` is the hash of the full-resolution render, not of the PNG here (see
"How the frames were made").
- **`SHA256SUMS`:** hashes of the PNGs in this repo; check with `sha256sum -c SHA256SUMS` inside each folder.
Class ids:
| Class | `metadata.jsonl` index | COCO `category_id` |
|---|---|---|
| streaks | 0 | 1 |
| dot_skip | 1 | 2 |
| haze | 2 | 3 |
| mottling | 3 | 4 |
| tunneling | 4 | 5 |
| bubbles | 5 | 6 |
| silvering | 6 | 7 |
| wrinkles | 7 | 8 |
### Loading
```python
from datasets import ClassLabel, List, Value, load_dataset
names = ["streaks", "dot_skip", "haze", "mottling", "tunneling", "bubbles", "silvering", "wrinkles"]
a = load_dataset("martinsawojide/webcheck-test-frames", "A", split="test")
# the loader infers plain integers from metadata.jsonl; cast to get the class names
a = a.cast_column("objects", {"bbox": List(List(Value("float32"), length=4)), "categories": List(ClassLabel(names=names))})
```
For the demo, download into a folder named `frames` at the root of the webcheck checkout:
```bash
hf download martinsawojide/webcheck-test-frames --repo-type dataset --local-dir frames
```
## How the frames were made
### Rendering
Rendered with BlenderProc 2.8.0 on its pinned Blender 4.2.1 (Cycles), from the scene scripts in the webcheck
repository (`scripts/render_web_a.py`, `scripts/render_web_b.py`). Each frame is rendered at 5376x3024 (0.18 mm per
pixel) and area-averaged to the 1920x1080 grid the labels use. The only image inputs are the project's own fictional
artwork and texture maps (`design/SKU_01/`); no third-party textures or HDRIs are used.
### Annotations
The boxes are not hand-drawn. The defect layout is planned in code before rendering; each box is that defect's
planned extent, cropped to the frame, then written to COCO. No person annotated the frames.
### Personal and sensitive information
None. The frames show a fictional product rendered from code; there are no people and no real brands.
## Limitations
- Synthetic frames of one fictional product and two film structures. Real presses differ in lighting, motion blur,
sensor noise, inks and substrates.
- Test split only.
- Haze and wrinkles are labelled but not covered by the demo models.
## Attribution
If you share or adapt these frames, CC BY 4.0 asks for credit. Suggested line:
> "webcheck synthetic print-defect test frames" © 2026 Martins Awojide, licensed under
> [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/), provided as-is without warranties.
> Source: https://huggingface.co/datasets/martinsawojide/webcheck-test-frames
If you change the frames, say so.
## Licence
Frames, labels and metadata: CC BY 4.0. The webcheck code that rendered them is MIT, except the Blender scripts,
which are GPL-3.0-or-later; the GPL does not cover the images they produce.