评测其他数据集的sh脚本？

[hshc123]

作者您好，我想在论文中提到的refcoco数据集进行eval，但是代码库里只有bash evaluation_scripts/eval_reasonseg.sh，请问后面会提供在其他数据的评测脚本嘛

[LiuRicky]

Replacing TEST_DATA_PATH="Ricky06662/ReasonSeg_test" to Ricky06662/refcocoplus_testA, Ricky06662/refcoco_testA or Ricky06662/refcocog_test. We currently provide these datasets. For others, you can generate dataset by yourself following the format of the above dataset.

[hshc123]

Replacing TEST_DATA_PATH="Ricky06662/ReasonSeg_test" to Ricky06662/refcocoplus_testA, Ricky06662/refcoco_testA or Ricky06662/refcocog_test. We currently provide these datasets. For others, you can generate dataset by yourself following the format of the above dataset.

感谢您的回复，还有一个困惑，就是我下载的refcoco和reason_seg数据文件夹格式是不一样的，reason_seg是
ReasonSeg_test
｜xxx.jpg,xxx.json

refcoco是
├── refcoco/
│ ├── refs(unc).p # pickle文件，包含引用信息
│ └── instances.json # COCO格式的标注文件
└── images/
└── mscoco/
└── images/
└── train2014/
└── COCO_train2014_*.jpg # 实际图像文件

[LiuRicky]

关于对refcoco数据的处理，可以参考https://github.com/dvlab-research/LISA/blob/main/utils/refer_seg_dataset.py ，你需要对应修改create_reasonseg_hf.py中数据读取的部分。
我们也会在接下来几周内陆续release refcoco(+/g)到huggingface format。