freek23/imagenet-enriched-grounded-boxes

ImageNet-enriched grounded boxes

Captions and grounded bounding boxes for the ImageNet-1k train split.

Images are not included. Each annotation is keyed by its ImageNet train image id (e.g. n13133613_29204); pair the annotations with an ImageNet-1k copy or with visual-layer/imagenet-1k-vl-enriched.

Contents

1,273 JSON shards covering 1,277,474 captioned samples, of which 801,472 carry grounded boxes (1,040,313 boxes in total). Boxes are filtered at grounding confidence ≥ 0.65 and shorter box side ≥ 32 px; samples whose candidate boxes all failed the filter keep their caption with an empty box list.

Provenance

Starting from the BLIP-2 captions of visual-layer/imagenet-1k-vl-enriched, boxes come from a two-stage pipeline:

1. Entity extraction — captions are parsed with spaCy; noun chunks survive a stop-list of ~110 abstract, meta-captional, placeholder, and locative root lemmas, yielding determiner-stripped entities and referring expressions.
2. Grounding — Grounding DINO T grounds all entities of an image in a single period-separated prompt; each returned (box, span, score) triple is matched back to its source chunk by an entity → root-lemma → substring cascade.

Format

Each shard train/train-sXXX-XXXXX.json maps image id → annotation:

{
  "n13133613_29204": {
    "caption": "a child eating corn on the cob",
    "boxes": [
      {
        "bbox": [199, 54, 389, 373],
        "confidence": 0.867,
        "entity": "child",
        "root": "child",
        "chunk_text": "a child",
        "referring_expression": "",
        "start": 0,
        "end": 7,
        "matched_label": "child"
      },
      {
        "bbox": [236, 185, 322, 232],
        "confidence": 0.827,
        "entity": "corn",
        "root": "corn",
        "chunk_text": "corn",
        "referring_expression": "corn on the cob",
        "start": 15,
        "end": 19,
        "matched_label": "corn"
      }
    ]
  }
}

bbox is [x0, y0, x1, y1] in pixels of the original image; start/end index the chunk's character span in the caption; referring_expression is the extended phrase when the chunk carries one; matched_label is the entity resolved by the match-back cascade.

Usage

To pack the annotations with their images into webdataset tars, see gromov-vlm, whose packer joins them with visual-layer/imagenet-1k-vl-enriched and whose filter defaults match the filtering above:

python datasets_prep/prepare_imagenet_enriched_webdataset.py \
    --annotations_dir path/to/train \
    --output_tar_directory datasets/train/imagenet_enriched