locate-anything
LocateAnything-Data
LocateAnything-Data
中文 · Paper ·
Model ·
Code
Overview
LocateAnything-Data is the public training-data release for
LocateAnything: Fast and High-Quality Vision-Language Grounding with Parallel
Box Decoding.
LocateAnything formulates detection and visual grounding as a unified
vision-language task. Given an image and a category, phrase, text string, or
action-oriented instruction, the model predicts the corresponding bounding
boxes or points. The data spans natural… See the full description on the dataset page: https://huggingface.co/datasets/NVEagle/LocateAnything-Data.LocateAnything-Data-ShareGPT-Annotationlocateanything-job-demo-warehouseLocateAnything-WDS
LocateAnything-WDS
由 NVEagle/LocateAnything-Data 转换的统一接口 WebDataset。
每个来源一个 config，命名 <domain>-<source>；样本粒度为 (view, 图像) 聚合。
样本结构（共享外层）
每个样本包含 __key__、图像成员（jpg/png/webp 等；annotation_only 来源无图像成员）
和 json。所有 domain 共享同一外层结构，差异仅在 role / task_type 取值域与 text 语义，
详见下方"数据 Schema 详解"。坐标为 0.0–1.0 归一化浮点：像素坐标 = x * width, y * height
（源为 0–1000 整型网格，除以 1000 转换，量化分辨率 1/1000）。
task_type = detection_grounding 时目标在 boxes，pointing 时在 points，
detection_grounding_pointing 时两者并存。写入侧… See the full description on the dataset page: https://huggingface.co/datasets/minato-ryan/LocateAnything-WDS.LocateAnything_datasetslocateanything-job-demo-hard-slow