RefStuff

This release contains two COCO 2014-derived splits for visual grounding with reference-style horizontal rows of repeated non-person objects.

Layout

• coco2014_candidates_4to8_unfiltered/: 15,577 candidate images. Each has one selected non-person COCO category with 4-8 annotated instances.
• strict_ready_131/: 131 visually screened samples with complete grounding annotations. Every item passed whole-image row screening and individual crop validation.

Strict grounding format

strict_ready_131/grounding.jsonl is the final annotation file. Each record includes the source image, target category, COCO instance IDs, visual row assignment, crop checks, and per-instance Chinese reference text. Reference text follows this form:

从上到下第N行从左数第K个<category>

or:

从上到下第N行从右数第K个<category>

Rows are numbered from top to bottom. The left/right ordinal is selected for each instance according to its side of the assigned horizontal row.

Provenance

Images and instance metadata originate from COCO 2014. The unfiltered split is a candidate pool, not a final-quality row dataset. The strict split is the screened subset intended for grounding use.