Why this matters LocateAnything‑Data aggregates the diverse spatial supervision needed to train unified vision–language grounding models: instead of juggling separate detection, grounding, OCR and GUI datasets, you get a single indexed repository of records (JSONL) and packed image shards plus mappings back to original media. That makes large‑scale, multi‑domain grounding experiments and distributed training workflows far easier to orchestrate.
What Sets It Apart
- Unified record format and normalized coordinates — All annotations use a shared
queryrepresentation and a 1000‑grid normalized spatial scheme, enabling consistent conversion to pixel coordinates across datasets so models can be trained on mixed tasks (bounding boxes and point targets) without per‑dataset reprojection logic. - Indexed WebDataset TAR shards + Energon metadata — Images are stored in indexed, compressed TAR shards for fast random access; Megatron‑Energon metadata and .tar.idx files let you plug the data directly into large‑scale distributed training pipelines.
- Rich, multi‑domain coverage — The release bundles views from many upstream sources (general detection, dense scenes, autonomous driving, GUI, OCR, documents/tables), plus mappings that let you trace every training record back to its original dataset filename and hash.
- Practical tooling and subset helpers — Includes download helpers, JSONL indexing utilities, example Visualizers and an Energon reader so you can inspect records, preview samples, or download curated subsets without fetching the entire multi‑terabyte payload.
Who It's For and Trade‑offs
Great fit if you: need a single, consistent spatial supervision corpus for vision–language grounding or pointing tasks; plan to run distributed training with Megatron‑Energon; or want to benchmark cross‑domain grounding models. Look elsewhere if you: require a single permissive license for all media (this collection preserves each source dataset's original licenses and some image pools must be hydrated from upstream), or you need fully reproducible Unsplash media (the release notes an unresolved Unsplash hydration path). The dataset prioritizes practical training plumbing and provenance over repackaging media under a unified license, so users must follow upstream usage terms and hydrate restricted pools as documented.
