MobileNet-YOLO
MobileNet-YOLO copied to clipboard
eric612
训练mAP很高(接近0.99),但是实际测试图片发现效果很差
我发现我的一个检测任务mAP和图片测试结果不吻合，经过排查不是过拟合的问题,因为我发现即使用那些验证集上的图片来进行测试，有些目标也没有被检测出来。同样的情况我更换另一个检测任务发现和mAP和测试结果符合。请问你知道是哪里出了问题吗
截图中红色椭圆型的是未检测出来的目标，按理说不应该出现这种情况才对呀
检查一下你的 Yolov3DetectionOutput layer 有没有设置confidence_threshold这个参数，不设置的话，默认是0.01
如果你test里面没有设置（默认0.01），而预测的时候用的0.3或者0.5，那么差距就会很大，可以将这两个值设置成一致的试试
检查一下你的 Yolov3DetectionOutput layer 有没有设置confidence_threshold这个参数，不设置的话，默认是0.01
如果你test里面没有设置（默认0.01），而预测的时候用的0.3或者0.5，那么差距就会很大，可以将这两个值设置成一致的试试
和conf_thres没关系,我已经放弃这个方案了
@pangkun248 How to check the precision and recall for every class?
@pangkun248 How to check the precision and recall for every class?
In this code, only the results of mAP and AP are given, but I think you can look at the source code. I think that since AP has been calculated, then recall and precision should be calculated. But I'm not familiar with C ++ and can only help you here.
Ok Thank you so much.
On Fri, Jan 3, 2020 at 3:17 PM pangkun248 [email protected] wrote:
@pangkun248 https://github.com/pangkun248 How to check the precision and recall for every class?
In this code, only the results of mAP and AP are given, but I think you can look at the source code. I think that since AP has been calculated, then recall and precision should be calculated. But I'm not familiar with C ++ and can only help you here.
— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/eric612/MobileNet-YOLO/issues/214?email_source=notifications&email_token=AG4GR5FDMW53YCME2PXENPDQ33RCBA5CNFSM4JRDSMCKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEIAPBGI#issuecomment-570486937, or unsubscribe https://github.com/notifications/unsubscribe-auth/AG4GR5CBZKJJ4NQWKGBQX2TQ33RCBANCNFSM4JRDSMCA .
test和train不一样的问题，应该主要是有域BN层引起的，你在test时将BN层的global参数设置成false结果应该可以，主要是过拟合或者训练样本不充分。
