README.md

---
license: cc-by-4.0
task_categories:
- image-classification
tags:
- 3d-printing
- fdm
- print-failure
- snapmaker-u1
size_categories:
- n<1K
pretty_name: 3d-cli print-failure frames
---

# 3d-cli print-failure frames

Camera frames from two FDM printers (Snapmaker U1, Creality K1 SE), labelled for the tier (b)
whole-frame vision model of [3d-cli](https://github.com/alex-mextner/3d-cli) (`3d printer vision`;
code: `tools/vision_train`). Classes: `healthy`, `spaghetti`, `part_detached`, `empty_bed`.

**This set is small and unbalanced.** 421 healthy frames from printing and a handful of
failure frames (7 empty_bed, 1 part_detached, 7 spaghetti). It is meant to be extended with frames recorded by `3d printer vision record`; it is
not enough to train a detachment detector on its own. `part_detached` is one event (a large
honeycomb panel that peeled off the plate): do not read a metric computed on it as evidence.

## Files

- `own/frames/<id>.jpg`: sanitised frames (`<id>` = first 16 hex of the file's sha256).
- `labels.jsonl`: `id, file, label, split (train|test), group (one print or event: keep a group
  on one side of a split), printer, source, width, height, sha256, qr_masked`.
- `sanitation.json`: per source how many frames were found, published, refused or masked, and the
  tool versions. `DATASET_SHA256`: sha256 over the sorted `id|sha256|label|split` lines of the
  published rows (the id of the dataset revision a model card pins).

## Sanitation

Frames were decoded and re-encoded (EXIF/ICC removed), cropped to the printer's bed-and-part area
(rooms, windows, screens and the chamber's back wall with its QR stickers are outside the crop),
QR codes were painted over, and a face and person detector (OpenCV Haar + HOG) refused any frame
with a hit (36 frames refused). File names carry no timestamps or project names. The
automatic checks are no proof: contact sheets of every published frame were reviewed by a human
before publishing. If you find a frame that should not be here, open an issue on the GitHub repo.

## Licence and provenance

All frames are the owner's own captures: CC-BY-4.0. The prints on the bed are the owner's own
models and models downloaded from model sites; no model files are included. The public frame
collections used for evaluation next to this set (spaghetti, bed-adhesion failures, ...) carry
their own licences and are **not** redistributed here; their provenance is documented in the
3d-cli repository (docs/research/printer-vision-tuning.md).

## Revision

Dataset sha256 `306bd93d73a5dfc7d966dfb6e230b2247ee6da5c42a6ecf0f62213b37791fc8e`; built by `tools/vision_train/dataset.py` at 3d-cli commit `a06ced7554e6d172844166675447f3e2d26c7e8a`.