Hi all,

While going through the training images, I noticed that most multi-line images are labeled with only the text of the centered/main line, but a portion of the training set has labels that include the full multi-line text instead.

Before I invest time building special handling for multi-line cases, I'd like to clarify: for the test set, are multi-line images expected to follow the same mixed convention (i.e., some evaluated as center-line-only, some as full multi-line), or is there a single consistent rule that applies to test specifically?

Concretely:

• Are all multi-line test images scored against just the center line's text?
• Or could some test images require the full multi-line transcription, matching how a subset of the training labels are structured?
• If it's a mix, is there any way to anticipate which case applies to a given image (e.g., based on how the image was cropped), or should we treat it as ambiguous and design for both?

Any clarification would help before I decide how much engineering effort to put into detecting/handling multi-line cases specifically. Thanks!