We're in the home stretch for ABC Project.
The pipeline is complete and we're currently debugging the final part (handling discards with cross-referenced indices).
The good news is that we've set aside EasyOCR in favor of a hybrid solution based on PP-OCR available in the Voyager-SDK: the flow involves running the model on Metis to identify all blocks on the covers, while a more advanced model running on CPU captures text with greater accuracy within the identified blocks.
The pipeline starts by using YOLOv8 to recognize the placement of a book in the loading area; at that point, the OCR acquisition is triggered (which uses various tricks to identify and detect text with increasing precision) and ultimately tries to match against the local database we've built.
After recognition, the system waits to determine whether the acquisition was correct (in which case you just need to move the book to confirm everything is fine, and it waits for the next one) or not — fingers crossed. We're working on this last part, which unfortunately generates false positives, but we're confident we'll be able to resolve it in the next few hours!
All updates in the official repo!
