search

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

multi_class_nms1(multi_bboxes, multi_scores, ......) num_classes = 81 ,but 81 ！= multi_bboxes.shape[1] #14

Open xcaizewu opened 3 years ago

xcaizewu commented 3 years ago

Hello, thank you for opening up your code

I am trying to port your code to mmdetection==2.3.0, I have a few questions： 1）When I trained the model, loss convergence was normal, but recall and AP were very low

WechatIMG66666

What part of the transplant was wrong with me

2）The method multi_class_nms1(multi_bboxes, multi_scores, ......）

WechatIMG333333

max_per_img= 125. Arguably, multi_bboxes.shape[1]=len(multi_scores)=125. But

WechatIMG2222222

multi_scores = 81, and for I in range(1, num_classes): multi_bboxes.shape[1]=125, It's not a complete cycle

JialeCao001 commented 3 years ago

@xcaizewu Thanks for your interest. I upload a new version using mmdetection v2.1.0. I think that the difference between v2.1.0 and v2.3.0 is not much. You can refer to this code or move the related files to your project. The link is https://github.com/JialeCao001/D2Det-mmdet2.1.

xcaizewu commented 3 years ago

@xcaizewu Thanks for your interest. I upload a new version using mmdetection v2.1.0. I think that the difference between v2.1.0 and v2.3.0 is not much. You can refer to this code or move the related files to your project. The link is https://github.com/JialeCao001/D2Det-mmdet2.1.

Thank you very much！ @JialeCao001