pranitchawla/FineGrainOCRITClustering

FineGrainOCR image-text clustering

This dataset adapts the complete usable validation split of FineGrainOCR for cross-modal clustering in MTEB. Each test row contains a grocery-product image, OCR derived from that image, and a product-class label registered by the checkout barcode scanner. Of the 18,416 source validation pairs, 27 rows with empty OCR are excluded, leaving 18,389 rows across all 256 product classes.

Benchmark input and output

The embedding model receives both image and text for each row and produces one vector. The clustering evaluator fits MiniBatch K-means to all vectors using the known number of classes (256). It does not give product labels to the model or clustering algorithm. The predicted cluster assignments are compared with label using V-measure.

The sample_id column is provenance metadata and is not model input.

Construction

Every validation pair with non-empty OCR is retained. Images are resized so their longest edge is at most 512 pixels and encoded as optimized JPEG at quality 90. Source members are decompressed and verified against the official ZIP's size and CRC-32 metadata.

OCR digit sequences containing 8–14 digits, including sequences separated by spaces or hyphens, are replaced by [BARCODE]. This prevents printed GTINs from revealing the class identifier while retaining product names and other package text.

The construction script and its focused tests are preserved on the codex/finegrainocr-build-scripts branch.

Provenance and license

• Source repository: https://github.com/Tubbias/finegrainocr
• Source commit: 9ce19719123fd33a994b103b6e91c37a640ce92b
• Paper: https://doi.org/10.1007/s00138-024-01549-9
• Source license: CC0-1.0

Build summary

{
  "source_commit": "9ce19719123fd33a994b103b6e91c37a640ce92b",
  "rows": 18389,
  "classes": 256,
  "min_rows_per_class": 6,
  "max_rows_per_class": 147,
  "barcode_redacted_rows": 5130,
  "barcode_redacted_sequences": 6725,
  "text_characters": {
    "min": 3,
    "median": 256,
    "p95": 1437,
    "max": 3628
  },
  "exact_duplicate_text_rows": 95,
  "cross_class_exact_duplicate_text_groups": 11,
  "image_max_edge": 512,
  "jpeg_quality": 90,
  "processed_image_bytes": 292796337,
  "image_width": {
    "min": 512,
    "median": 512,
    "max": 512
  },
  "image_height": {
    "min": 384,
    "median": 384,
    "max": 384
  },
  "source_compressed_bytes": 10664221585,
  "coalesced_image_transfer_bytes": 10636823036,
  "image_archive_ranges": 160
}

Citation

@article{pettersson2024,
  title = {Multimodal fine-grained grocery product recognition using image and OCR text},
  author = {Pettersson, Tobias and Riveiro, Maria and L{\"o}fstr{\"o}m, Tuwe},
  journal = {Machine Vision and Applications},
  volume = {35},
  number = {4},
  pages = {79},
  year = {2024},
  publisher = {Springer},
  doi = {10.1007/s00138-024-01549-9},
}