uv-scripts/vlm-object-detection
收藏资源简介:
VLM目标检测数据集是一个用于通过视觉语言模型(VLM)进行指令提示目标检测的资源,使用vLLM实现。它被设计为VLM作为标注器的原语,用于引导目标检测数据集的创建——用户可以提供自由形式的提示(如检测每张照片和插图、检测所有个人防护装备物品、检测每个电子元件并识别其参考标识符),模型将返回边界框JSON,适用于下游标注工具(如Label Studio、FiftyOne、COCO转换)。该数据集包括脚本和示例,支持零样本和少样本检测,适用于自然图像、文档布局等多种场景,但输出为标注级而非训练级,主要用于标注管道的初始种子。
The VLM Object Detection dataset is a resource for instruction-prompted object detection with vision-language models (VLM) via vLLM. It is designed as a VLM-as-labeller primitive for bootstrapping object-detection datasets—users can provide free-form prompts (e.g., detect every photograph and illustration, detect all PPE items, detect every electronic component and identify its reference designator), and the model returns bbox JSON ready for downstream labelling tools (Label Studio, FiftyOne, COCO conversion). The dataset includes scripts and examples, supporting zero-shot and few-shot detection for various scenarios such as natural images and document layouts, but the output is labelling-grade not training-grade, primarily serving as an initial seed for labelling pipelines.




