Hi — a rules clarification, following the helpful answers in #33870, #33891 and #34459.

Concretely:

1. Segment character-level crops from the provided Train.csv images only.

2. Recombine those crops side by side to render new line images, paired with the text they spell.

3. Add the result to training as augmentation, alongside the real lines.

To be precise about what this is not:

- No external dataset. Every pixel comes from the challenge's own training images.

- No generative model. It is cut-and-paste of existing crops, not synthesis by a GAN, diffusion model, or font renderer.

- No test data involved. Test images are not used in any way here.

- No manual annotation. The whole pipeline is code; the label of a generated line is the concatenation of the labels of the crops it is built from.

My reading is that this is closer to standard image augmentation (already common practice, and I already use geometric and photometric augmentation) than to "using an external dataset". But it does create new training images, so I would rather ask than assume.

Question: is this permitted?

Thank you all.