ChrisRPL/blackline-atlas-sam3-real-eval-v2

该数据集名为Blackline Atlas SAM3真实图像评估包，是一个评估和集成数据集，而非训练基准。它包含22个案例，44个PNG图像文件，以及一个JSONL格式的数据集文件。图像布局包括当前和基线卫星帧。数据集的主要用途是在决定是否需要分割微调之前，对选定的民用中断站点运行SAM3/SAM3.1零样本可提示分割。数据集的重要限制包括它不是战术目标数据集，标签是Blackline操作评估标签，而非专家像素掩码，以及Sentinel分辨率限制了小对象和细粒度损伤声明。

This dataset, named Blackline Atlas SAM3 Real-Image Eval Pack, is an evaluation and integration dataset, not a training benchmark. It includes 22 cases, 44 PNG image files, and a JSONL format dataset file. The image layout consists of current and baseline satellite frames. The primary purpose of the dataset is to run SAM3/SAM3.1 zero-shot promptable segmentation against selected civilian disruption sites before deciding whether any segmentation fine-tuning is justified. Important limitations of the dataset include that it is not a tactical targeting dataset, the labels are Blackline operational eval labels and not expert pixel masks, and Sentinel resolution limits small-object and fine-grained damage claims.