The Swiss-French SignPuddle includes around 5000 entries in SignWriting with an illustration.
We collected the data using collect_data.py.
Since this data is unstructured, we assume all png files are illustrations, and all jpg files are pictographs.
This means that we have multiple types of data:
| Type | Example |
|---|---|
| Noise |  |
| No-person Illustration |  |
| Blurry Illustration |  |
| HD Illustration |  |
| HD Colored Illustration |  |
| Corrupted |  |
| Blurry Image + Arrows |  |
| HD Image |  |
Ideally, we should be able to classify these images, and apply some "tag" in their prompt. We should not train on corrupted or noise images.