LiWentomng / Point2Mask

The code for "Point2Mask: Point-supervised Panoptic Segmentation via Optimal Transport", ICCV2023
Apache License 2.0
65 stars 6 forks source link

模糊图上的效果 #1

Open tongchangD opened 1 year ago

tongchangD commented 1 year ago

请问有在模糊数据上试过这种方案的效果吗 我发现现在分割比较好的展示都是在边缘还算比较清晰的数据上,而真实落地时会出现许多模糊的图像,这种图像边界不够清晰,会不会造成误分割

LiWentomng commented 1 year ago

@tongchangD 对于模糊的数据 可能work的不好,算法本身依靠图像low-level的信息,包括color、pixel差异做为基础,如果模糊数据导致目标之间差异较小,导致效果不好。 此外,对于密集的同类目标,该算法表现也不佳,确实存在比较大的挑战。 这些也是未来这个task需要重点解决的问题。

tongchangD commented 1 year ago

@LiWentomng 谢谢你的回复 祝好, tongchangD

tongchangD commented 1 year ago

@LiWentomng 训练数据这块能否提供由几张图构成的简要demo,你现在提供的大数据集,忒难下载了。

LiWentomng commented 1 year ago

@tongchangD 可以优先考虑 VOC数据集,该数据集为相对小数据集,大约有1.29G,下载链接为Google Driver, 对应point 标注链接为Google Drive.

tongchangD commented 1 year ago

@LiWentomng 非常感谢,我试试