JialeCao001 / D2Det

D2Det: Towards High Quality Object Detection and Instance Segmentation (CVPR2020)
https://openaccess.thecvf.com/content_CVPR_2020/papers/Cao_D2Det_Towards_High_Quality_Object_Detection_and_Instance_Segmentation_CVPR_2020_paper.pdf
MIT License
297 stars 86 forks source link

Excuse me, how can I realize the training of D2Det model with iSAID data set #33

Open zhaozhiyi11 opened 3 years ago

zhaozhiyi11 commented 3 years ago

Excuse me, how can I realize the training of D2Det model with iSAID data set

JialeCao001 commented 3 years ago

@zhaozhiyi11 Thanks for interests. I remeber that some parameters should be changed (e.g., the number of proposals), which follow the settings reported in iSAID paper.

zhaozhiyi11 commented 3 years ago

@zhaozhiyi11 Thanks for interests. I remeber that some parameters should be changed (e.g., the number of proposals), which follow the settings reported in iSAID paper.

Thanks for your reply.Have you ever trained D2Det with an iSAID dataset?Can you share some of the models you've trained with the iSAID dataset?proposals are put in which document?

JialeCao001 commented 3 years ago

@zhaozhiyi11 I am sorry that I donot save the model on iSAID from the foreign server. The parameter settings follow the iSAID paper https://openaccess.thecvf.com/content_CVPRW_2019/papers/DOAI/Zamir_iSAID_A_Large-scale_Dataset_for_Instance_Segmentation_in_Aerial_Images_CVPRW_2019_paper.pdf.

zhaozhiyi11 commented 3 years ago

@zhaozhiyi11 I am sorry that I donot save the model on iSAID from the foreign server. The parameter settings follow the iSAID paper https://openaccess.thecvf.com/content_CVPRW_2019/papers/DOAI/Zamir_iSAID_A_Large-scale_Dataset_for_Instance_Segmentation_in_Aerial_Images_CVPRW_2019_paper.pdf.

谢谢您的回复!这边论文我已经看过了,您说的修改参数是在D2Det的模型的config里修改吗?

JialeCao001 commented 3 years ago

对了,需要把一些coco的参数按照这个论文第7页里面的说法进行修改。

zhaozhiyi11 commented 3 years ago

对了,需要把一些coco的参数按照这个论文第7页里面的说法进行修改。 您好!我还是没有找到在要修改哪些参数,您能具体说一下吗?麻烦您了!

JialeCao001 commented 3 years ago

@zhaozhiyi11 主要可能是下面这些参数: train_cfg: rcnn.sampler.num = 1024, rcnn.sampler.pos_fraction= 0.5;test_cfg:rcnn.max_per_img=500; test_pipeline: img_scale=[(1333, 800),(1333, 1200),(1333, 1000),(1333, 600),(1333, 400)]. 如果还有啥调参的问题也可以加我微信。