sensenova/SenseNova-Vision-Benchmark
收藏资源简介：
SenseNova Vision Benchmark JSONL 是一个用于评估开源视觉语言模型的基准数据集，专注于检测和分割任务。该数据集包含52个JSONL文件，覆盖对象检测、图像分割、OCR（光学字符识别）、文档理解、接地（grounding）、关键点检测、点接地（point grounding）、视觉提示（visual prompt）以及GUI（图形用户界面）图标和文本定位等多种计算机视觉任务。具体任务包括：COCO、LVIS、VisDrone等标准检测数据集；DocLayNet、HierText等文档分析任务；RefCOCOg等引用表达式理解；以及ADE20K全景分割、COCO全景分割、GCG分割、推理分割和二进制分割等分割任务。数据集以JSONL格式组织，便于直接集成到统一评估脚本或转换为其他评估框架所需格式。所有文件分为检测和分割两大目录，检测目录下进一步细分为GUI评估、关键点评估、点评估和视觉提示评估子目录。该数据集旨在为多模态生成模型提供全面的视觉任务评估基准。
SenseNova Vision Benchmark JSONL is a benchmark dataset for evaluating open-source vision-language models, covering detection and segmentation tasks. It consists of 52 JSONL files that encompass a wide range of computer vision tasks, including object detection, image segmentation, OCR (optical character recognition), document understanding, grounding, keypoint detection, point grounding, visual prompting, and GUI (graphical user interface) icon and text localization. Specific tasks involve standard detection datasets such as COCO, LVIS, and VisDrone; document analysis tasks like DocLayNet and HierText; referring expression comprehension such as RefCOCOg; and segmentation tasks including ADE20K panoptic segmentation, COCO panoptic segmentation, GCG segmentation, reasoning segmentation, and binary segmentation. The dataset is organized in JSONL format, making it easy to plug into unified evaluation scripts or convert to formats required by other evaluation frameworks. All files are divided into two main directories: detection and segmentation, with the detection directory further subdivided into GUI evaluation, keypoint evaluation, point evaluation, and visual prompt evaluation subdirectories. This dataset aims to provide a comprehensive benchmark for evaluating multimodal generation models on diverse vision tasks.
