Open befunger opened 9 months ago
Attempting to work with the limitation by using an LLM to extract unique book titles from the complete text extracted from a single detection. Some promising results but also a lot of garbage. Might be better to just fix the issue at its core.
Bounding boxes end up capturing not just a single book spine, but a lot of adjacent ones too, which messes with the text recognition and leads to mixed and duplicate text in output. Forced square shape of bounding boxes does not help here either.