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

关于关键点评估的请教 #6

Open
DaMiBear opened this issue Jun 20, 2022 · 6 comments
Open

关于关键点评估的请教 #6

DaMiBear opened this issue Jun 20, 2022 · 6 comments

Comments

@DaMiBear
Copy link

您好,我看了您的程序,发现您是把关键点的评估注释掉了是吗?
我看了Ti的那个yolo-pose的程序,好像也没有对关键点的评估。
那请问您是基于什么原因去掉了关键点的评估呢?

@mcmingchang
Copy link
Owner

哈哈,谢谢关注
主要是平时也没有怎么维护,训练出来的结果肉眼看上去比之前提升不少也更稳定就没有去写这个啦,毕竟不是正规的项目,仅做分享,而且我这做项目的也没有那么多测试集要跑,所以就没写啦

@mcmingchang
Copy link
Owner

这个应该也不难,有很多关键点的项目可以参考啦,粗暴点可以全部mse来算

@DaMiBear
Copy link
Author

嗯是的,原本是看了您和Ti的代码,感觉都没有用关键点评估,都怀疑是pycocotools的关键点评估的问题了😂
我看coco的pycocotools里是可以设置iouTypekeypoint的,计算的是OKS。目前只用了coco2017的人体关键点的val数据集测试了一下,改为keyoints后评估输出的结果和iou的格式是一样的,但全是0,OKS loss在训练过程中还是有所下降的3.3->1.4,测试了几张图片关键点能看到一点点效果,所以才怀疑自己的用法是不是有问题;也可能是我网络太轻量化的原因,因为刚刚接触关键点这一块,这个我还要再研究怎么使用pycocotools的关键点评估。

Average forward time: 0.43 ms, Average NMS time: 0.74 ms, Average inference time: 1.17 ms
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets= 20 ] = 0.000
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets= 20 ] = 0.000
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets= 20 ] = 0.000
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets= 20 ] = 0.000
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets= 20 ] = 0.000
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 20 ] = 0.000
 Average Recall     (AR) @[ IoU=0.50      | area=   all | maxDets= 20 ] = 0.000
 Average Recall     (AR) @[ IoU=0.75      | area=   all | maxDets= 20 ] = 0.000
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets= 20 ] = 0.000
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets= 20 ] = 0.000

@mcmingchang
Copy link
Owner

这我就不清楚啦,反正我这用在实际项目上效果还不错

@mcmingchang
Copy link
Owner

你可以你看看head那里的代码,kps的头我改小了点,你可以按论文标准的来,准确率高一点,但是速度会降一点

@DaMiBear
Copy link
Author

我网络的backbone是shufflenetV2,就是想速度越快越好😂。
本来只有目标检测,现在想加一个手部的关键点检测的预测头,所以关键点部分参考了您的代码,至于结果都为0应该是我关键点评估的代码有问题,这个我研究研究应该能解决。

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

2 participants