foralliance
foralliance
HI Is it convenient to provide the **DSFD-ResNet50** model?? many many thanks!! my emial is [[email protected]](url)!!
from the code, i found that "multi-scale training" means : randomly select a scale from multiple scales. u only need to train once. is that right?? Because in the past,...
@bharatsingh430 @bharatpublic HI In README, you point out: `Please use the version of caffe matches with this repository. I have merged many files between the latest version of caffe and...
@bharatsingh430 @bharatpublic HI About the **"Results on ms-coco" section** Based on the training model you provided, 1. In **master branch**, the test results were **29.0** 2. In **coco branch**, the...
@bharatsingh430 HI Section 5 of the paper points out: `Simple modifications like 5 scales for RPN anchors, a minimum image size of 800, 16 images per minibatch and 256 ROIs...
@c0nn3r HI For focal loss. The **classification branch** uses the **sigmoid function**. Why **background class** is considered in the classification branch??such as **coco**, num_classes=80 instead of 81.
@c0nn3r HI what's your **python** version and **pytorch** version?
在你的code中，conv6的形式是： conv6 = nn.Conv2d(512, 1024, kernel_size=3, padding=6, dilation=6) 但这个形式其实是SSD中的。 而在PyramidBox中，底层是基于S3FD中的VGG16。跟SSD没关系。所以conv6的形式应该是： conv6 = nn.Conv2d(512, 1024, kernel_size=3, padding=1)
@yxlijun HI 看了您的code，受益匪浅 起初看paper时，因为公式较多，很多地方难以理解．结合您的code来看，很多都理解了．非常感谢！ **整体感觉：** 从code来看，对于**损失函数**这里，其实还是常规的．重点在于**head标签信息**的设置，其本质上是通过**face的标签信息**得到的： 1. 关于head坐标信息，体现在data/widerface.py中，head的坐标就是在face的基础上扩大了一倍．这个感觉就对应的是**公式(5)**的内容 2. 关于head类别标签（0/1）信息，和face标签保持一致． **在正负样本匹配时：** 从code来看，不论是face还是head，共用一组anchor（priors）． 比如一个face，尺度128．那么这个face对应的合理的anchor就会出现在第4个预测分支上．这个face对应head的尺度就是256，那么这个head对应的合理的anchor就会出现在第5个预测分支上． **有个疑问**：因为head是由这个face衍生来的，那么需不需要把这个2个anchor标记一下，表示其来自于同一个face呢？？从**公式(1),(4)**来看，face和head的anchor总感觉有一种匹配关系．但是从code来看，没有这样的标记，针对face和head的损失都是独立进行的． 另外有[另一个版本的PyramidBox复现代码](https://github.com/Goingqs/PyramidBox)，他针对face和head设计了2组anchors．但总感觉head_anchors设计的有问题．不知道您参考过没？
1节“Feature learning部分”指出： `现有的FPN方法just aggregates hierarchical feature maps between high and low-level output layers, which does not consider the current layers information, and the context relationship between anchors is ignored。` 2节“Feature...