EMD移植到yolo出现重复框

[Eagle104fred]

您好，请问一下，我把您的算法和思路移植到了yolo系列下，您的论文中对此的可行性也有提及，可是我在使用相同的crowdhuman数据集进行校验是出现了检测框重叠的问题，按照您的setNMS算法，检测框确实会出现重叠在同一目标的情况，请问您是怎么解决这一问题的，我在您的论文和开源的代码RetinaNet实现中没有发现相应的解决方案。

[Eagle104fred]

[Eagle104fred]

我想了一下，是否是我loss函数设置有问题，两个emd头应该是不能使用相同的gt的

[yexiguafuqihao]

Maybe it is one of the reasons. In our paper, a proposal should predict 2 detections to cover 2 different targets. And our schema is mainly deployed to the two-stage detection frameworks. It will aggravate the problem of imbalance between positive and negative samples because an anchor infers at least 2 detections in the one-stage detection framework. Therefore, the performance could be better if the imbalance problem can be addressed. Besides, it does not work when the focal loss is included in the EMD loss.

[Mobu59]

您好，请问一下，我把您的算法和思路移植到了yolo系列下，您的论文中对此的可行性也有提及，可是我在使用相同的crowdhuman数据集进行校验是出现了检测框重叠的问题，按照您的setNMS算法，检测框确实会出现重叠在同一目标的情况，请问您是怎么解决这一问题的，我在您的论文和开源的代码RetinaNet实现中没有发现相应的解决方案。

您好，能请问一下您是怎么将EMD移植到YOLO系列中的呢，求助！

[Eagle104fred]

失败了，收敛不了，还在研究 钟文华

…

在 2022年3月11日，11:23，Mobu59 ***@***.***> 写道：  您好，请问一下，我把您的算法和思路移植到了yolo系列下，您的论文中对此的可行性也有提及，可是我在使用相同的crowdhuman数据集进行校验是出现了检测框重叠的问题，按照您的setNMS算法，检测框确实会出现重叠在同一目标的情况，请问您是怎么解决这一问题的，我在您的论文和开源的代码RetinaNet实现中没有发现相应的解决方案。 您好，能请问一下您是怎么将EMD移植到YOLO系列中的呢，求助！ — Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android. You are receiving this because you authored the thread.

[yexiguafuqihao]

你可以尝试直接使用focal loss而不是emd loss,这样就能够收敛了,也就是说模型在训练过程中暂时不用先考虑permutation invariance.

[Eagle104fred]

单focalloss已经有人实现了呀，官方就自带这个方法 钟文华

…

在 2022年3月12日，16:42，yexiguafuqihao ***@***.***> 写道：  你可以尝试直接使用focal loss而不是emd loss,这样就能够收敛了,也就是说模型在训练过程中暂时不用先考虑permutation invariance. — Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android. You are receiving this because you authored the thread.

[yexiguafuqihao]

如果是做工程的话,可以尝试:https://arxiv.org/pdf/2012.03544.pdf.端到端的物体检测,做工程落地的话已经够了.

[InterstellarFang]

所以想问下EMDloss不能用在one-stage目标检测算法中吗

[InterstellarFang]

Maybe it is one of the reasons. In our paper, a proposal should predict 2 detections to cover 2 different targets. And our schema is mainly deployed to the two-stage detection frameworks. It will aggravate the problem of imbalance between positive and negative samples because an anchor infers at least 2 detections in the one-stage detection framework. Therefore, the performance could be better if the imbalance problem can be addressed. Besides, it does not work when the focal loss is included in the EMD loss.

所以想问下EMDloss不能用在one-stage目标检测算法中吗