lance-format/coco-detection-2017-lance

COCO 2017对象检测（Lance格式）数据集是COCO 2017对象检测基准的Lance格式版本，源自Hugging Face上的detection-datasets/coco数据集。该数据集以Lance格式存储，每行代表一张图像，包含内联的JPEG字节、完整的边界框列表、COCO 80类类别ID和名称、每个对象的面积、OpenCLIP图像嵌入以及预构建的索引。数据集支持快速随机访问、多模态存储和高效查询，适用于对象检测、图像检索和机器学习训练等任务。数据集分为训练集（117,000+行）和验证集（4,950+行），总标注框数约为860k（训练）和37k（验证）。关键特性包括内联图像字节、并行标注列、预聚合摘要和CLIP图像嵌入。此外，数据集还提供了预构建的索引，如IVF_PQ向量索引和BTREE索引，以支持相似性搜索和快速查找。数据集可以通过Hugging Face的datasets库、LanceDB或Lance直接加载，并支持数据演进、版本控制和子集材料化等高级功能。COCO 2017 object detection (Lance format) dataset is the Lance-formatted version of the COCO 2017 object detection benchmark, derived from the detection-datasets/coco dataset hosted on Hugging Face. Stored in Lance format, each row in this dataset represents an image, containing inline JPEG bytes, a complete list of bounding boxes, COCO 80-class category IDs and names, the area of each object, OpenCLIP image embeddings, and pre-built indexes. This dataset supports fast random access, multimodal storage and efficient querying, and is applicable to tasks including object detection, image retrieval and machine learning training. It is split into a training set (over 117,000 rows) and a validation set (over 4,950 rows), with approximately 860k annotation boxes in the training split and 37k in the validation split. Key features include inline image bytes, parallel annotation columns, pre-aggregated summaries and CLIP image embeddings. In addition, the dataset provides pre-built indexes such as IVF_PQ vector indexes and BTREE indexes to support similarity search and fast lookup. The dataset can be directly loaded via Hugging Face's datasets library, LanceDB or Lance, and supports advanced functionalities such as data evolution, version control and subset materialization.