Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

can it used in DOTA dataset? #1

Open
Smoothing97 opened this issue May 11, 2019 · 9 comments
Open

can it used in DOTA dataset? #1

Smoothing97 opened this issue May 11, 2019 · 9 comments

Comments

@Smoothing97
Copy link

Thanks for your work:)and I wanna know can I use it in DOTA dataset?Cause I see the code structure is similar with r2cnn.If it can be used in DOTA,where should I notice?I try to modify the cfgs.py and read_tfrecords.py,stiil using 800*800 cropped images.hope to know the author's opinion about using it in DOTA.
Hope for your reply,thank you!

@Smoothing97
Copy link
Author

update:using it training DOTA_H...wait the testing result:)

@yangxue0827
Copy link
Member

According to the experience of training coco, cascade rcnn does not perform as well as FPN in ap50. Looking forward to your results。 @Smoothing97

@yangxue0827
Copy link
Member

Recommend NAS_FPN_Tensorflow @Smoothing97

@Smoothing97
Copy link
Author

昨天我分别尝试了NAS_FPN和CASCADE_FPN。
为了节省重新生成DOTA_H_train.tfrecord的时间,我改写了读取tfrecord的代码,因为当前需要读入(xmin,xmax,ymin,ymax,label)的格式,而我的tfrecord是R2CNN中8dots+label的格式。
因为GPU数量有限,我没有采用multi_gpu_train,两份程序都是在一张1080TI上训练的,按照R2CNN的训练经验,我以为1天到两天就可以训练完毕,于是美滋滋的去睡觉了:)
今天早上查看结果,训练10小时后,NAS_FPN训练了2w步,cascade_FPN训练了3w步。ORZ
我能最快调动的1080TI最多有4张,所以NAS_FPN还是很吃显卡的对吗【手动狗头

@yangxue0827
Copy link
Member

nasfpn你可以在cfgs那选择配置文件,默认 @Smoothing97 的是stack七次,通道384,模型比较大

@yangxue0827
Copy link
Member

如果gpu有限,建议用multi gpu train一个一个模型训练,这样会快一点 @Smoothing97

@Smoothing97
Copy link
Author

@yangxue0827 经历了断网等一系列操作,我现在决定用可用的3张1080TI来训练cascade_fpn,数据集是DOTA_H。NAS_FPN比较费时间,先调一个比faster rcnn好的版本出来:)
另外,因为最近在研究DOTA数据集上的检测,基于Faster RCNN,我在想PANet会不会比FPN要好一些?是基于mask rcnn多做了一次特征融合。
image
我的思路还是跟进近年来rcnn的发展,引用自然图像中的检测算法,现在正在不断测试,也没有做数据增强和对于遥感图像的处理。因为看到大佬您在DOTA的检测上做了很多工作,想请教您在这个问题上的一些心得:加上FPN、PANet、NAS-FPN、ROI align这些模块对精度的提升作用大吗?还是应该设计更好的网络结构,或者采用one-stage的模型会有帮助吗?
打扰了,期待您的回复:)

@Smoothing97
Copy link
Author

update:
NAS-FPN似乎对小目标友好,stack七次,通道384下3块1080TI训练10h,step到25000。打开tensorboard看检测结果:
image
image
image
但是对大目标或者肉眼明显的目标还没显示出效果:
image
由于训练步数太短,我将num_mas从7降为3继续训练,不知道会产生什么结果。等待ing...

@zhang295498
Copy link

update:
NAS-FPN似乎对小目标友好,stack七次,通道384下3块1080TI训练10h,step到25000。打开tensorboard看检测结果:
image
image
image
但是对大目标或者肉眼明显的目标还没显示出效果:
image
由于训练步数太短,我将num_mas从7降为3继续训练,不知道会产生什么结果。等待ing...

您好,您的实验做的怎么样了?有什么进展么?

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

3 participants