elliot-mllm/localize-general-objects

Elliot Localize General Objects

Upstream training splits; known explicitly identified test/eval rows excluded. Cross-dataset benchmark overlap is not guaranteed. Published as a raw, manually gated release; source annotation caveats remain.

Task views reuse original image archives. Coordinates are normalized 0–1000. Within each query targets are sorted left-to-right then top-to-bottom.

HF preview configs contain 10 examples per view, not the complete training split. Full training uses metadataset.yaml and examples/read_energon.py (Energon 7.4).

Source	View	Records	Targets	Examples
blip3_grounding	boxes	7,319	51,243	10 annotated examples
grand	detection	3,819,625	189,778,638	10 annotated examples
grand	referring_expression	3,761,693	27,165,058	10 annotated examples
grand	affordance_grounding	2,151,574	28,936,993	10 annotated examples
grand	grounded_captioning (optional; excluded from default)	3,808,343	16,606,582	10 annotated examples
grand	grounded_captioning (optional; excluded from default)	7,536,855	16,094,385	10 annotated examples
grit	boxes	6,010	10,625	10 annotated examples
grit	boxes	6,010	10,625	10 annotated examples
lvis	boxes	99,388	1,270,141	10 annotated examples
pascal_voc	boxes	8,218	23,618	10 annotated examples
roboflow100	boxes	153,963	975,375	10 annotated examples

Annotation and image origins

Source	Upstream annotations	Image origin / reuse	Coverage
blip3_grounding	Salesforce BLIP-3 grounding	Web URLs; released example data_type=detection_datacomp. Full underlying mixture not independently verified.	7,319 recovered media rows from a 10,000-row recovery attempt; NOT full 50M release
grand	MBZUAI GranD / GLaMM	SA-1B images; GranD adds grounded annotations. Images shared with sa1b_boxes.	grand_rich_v2: five task-separated views; dense/short captions optional and excluded from default. Shared SA-1B images; counts overlap.
grit	Microsoft Kosmos-2 GRIT (not AllenAI GRIT benchmark)	COYO-700M / LAION-2B web image-text pairs; local source filenames coyo_*.parquet.	6,010 recovered images from a 10,000-row recovery attempt; two alternate text views sharing images and boxes. NOT 20.5M complete release.
lvis	LVIS v1	COCO images, additional long-tail LVIS labels; same image substrate as COCO/Osprey/other COCO-derived sources.	99,388 exported LVIS train rows; native boxes only, not polygons.
pascal_voc	PASCAL VOC; HuggingFaceM4 mirror	VOC images; reused by Pascal Part family. First export provenance is VOC2007 XML.	8,218 selected official-train records; not test/validation or complete upstream dataset.
roboflow100	Roboflow-100; Myszz mirror	100 community datasets across domains. Per-record source_annotation.dataset retains sub-dataset; no claim of globally unique images.	153,963 train_cleaned mirror records; fixed image mapping bundle. Not newer RF100-VL.

What is not claimed

No guarantee of cross-dataset benchmark disjointness, global image uniqueness, semantic annotation correctness, or redistribution/license clearance. Shared box/point views are alternate tasks, not distinct images. Source-specific caveats remain applicable.

The package preserves upstream query text; metadata cleanup does not alter images or embedded image metadata.

Image packaging — GranD annotations only

GranD includes all five annotation views, indexes, and small illustrated previews, but does not include the bulk SA-1B image archives. All other sources include their image archives. This release is therefore not fully self-contained.

• Use manifests/metadataset_bundled_images.yaml to load only sources whose images are included (BLIP-3 recovered subset, GRIT recovered subset, LVIS, VOC, Roboflow).
• Root metadataset.yaml includes GranD detection/referring/affordance and requires external SA-1B media. GranD dense/short captions remain optional.
• manifests/grand_external_media.json lists every omitted prepared shard and size. Existing GranD media indexes and configs are retained. They work with the exact prepared shards in that manifest, not arbitrary original SA-1B TARs.
• If supplying the exact prepared pools, place or symlink them at the manifest's relative paths under the downloaded release. Our existing local training package already has these files; no changes are required there.
• If supplying original/differently packed SA-1B archives, rebuild Energon media indexes for those archives and update each GranD view's aux pool paths while preserving annotation image member names. That reconstruction is not automated by this release; renaming archives will not fix byte offsets.

The record counts are task-view records, not globally unique images. Small preview images are intentionally included even for GranD. Cross-dataset leakage and image redistribution rights have not been globally verified.