Proactively let negative samples participate in training problems(主动让负样本参与训练问题)

[yiran-THU]

trafficstars

Hi guys,

In engineering projects, there are often situations in which a picture is completely unmarked. In the process of detecting, there are many misidentified results. Previously, in the darknet framework, directly adding some unlabeled information to the training data set can solve these problems. Now, I want to use caffe to train this model of yolov3, but I find it seems that I can't actively add images with only negative samples for training.

If you want to train by directly adding pure negative samples, how to deal with it, thank you~

---

在实地的工程项目中，经常会遇到一些情况，就是一个完全没有标记样本的图片，在进行检测的过程中，有很多被误识别的结果出来。

之前在darknet框架下的方式是，直接增加一些没有标注的信息的图片，增加到训练数据集中，可以去解决这些问题。 现在，我想用caffe去训练一下yolov3的这个模型，但是发现好像不能去主动添加完全只有负样本的图片进行训练。

如果想通过直接增加纯负样本的方式来进行训练，该怎么处理，谢谢~

---

@eric612 大佬 能否指点一下 谢谢

[eric612]

可試試這個 https://github.com/eric612/MobileNet-YOLO/issues/206

[yiran-THU]

好的，我试试，谢谢！

[yyqgood]

@yiran-THU 你好， 你使用以上方式，有降低误检吗？