数据
Opened this issue · 12 comments
数据在哪里下载呀
DOTA数据可以在这下载 https://captain-whu.github.io/DOTA/dataset.html
I'm learning your code about Oriented RepPoints. But I am a novice,there are many problems in data set preparation,I have just come into contact with this direction, so I am not very clear about the conversion and processing of data formats.
Can you send me a complete set of data you have prepared?(as on your web page)
@ziruoliu
Hello, the code for DOTA data preparation, please refer this code-prepare_dota1_ms.py to get single-scale image patches. The json format is similar as COCO json with 8 points(ploy) for box annotation as in DOTA2COCO_poly.py.
@LiWentomng Thank you very much for your reply, I have one more question for you. I was reading the paper to see experiments about angular errors. I don't see where it comes from, I wonder how you define it.
"mean Average Orientation Error (mAOE◦) "
Look forward to your reply
你好,请问在训练之前dota数据集需要进行切割吗?如果需要,请问切割成多大大小合适?
@LiWentomng 你好,还有一个问题想请教一下,dota数据集分割之后,有的图片里面就没有目标,它所对应的标注文件也是空的。那么在训练的时候,这些图片需要剔除掉吗?还是说数据集分割完之后,就不要去管这些问题了。
@runever 你好 不需要自己手动剔除, 本身mmdetection框架下对应的data pipeline 有filter 的操作,过滤掉没有标注的图片,本代码也有过滤掉特别小尺寸的目标的操作。
@LiWentomng 多谢多谢解惑
@LiWentomng 作者您好,我看您在论文中说对于dota数据集“We crop the original images into the patches of 1024 × 1024 with a stride of 824”,但项目文件prepare_dota1_ms.py中gap设置为了824,这样ImgSplit_multi_process.py中的步长slide=subsize-gap岂不是1024-824=200,这个地方不太明白,可能是我理解错了或是gap应该设置成200呢?