about train

[fengxiuyaun]

你好，非常感谢您的分享-IOUnet，但是我有几处不了解的地方。

1. 文章说For a better performance, we use class-aware IoU predictors.这个是什么意思呢? 2.在3.1 Learning to predict IoU节，这部分仅是介绍了怎么训练Standalone IoU-Net吧？就如rcnn一样分两步：第一步先训练前面的FPN。第二步，在您说的augment GT上训练IOU那个分支。再第二步的时候把augment的产生的box和第一步产生的feature输入到PrRoI-Pooling中，预测得到的IOU与IOU标签（取值[-1,1]）作loss，进行训练？ 3.假如联合训练的话就是再2中第二步添加下边的分类和回归分支？

[fengxiuyaun]

还有就是，我还是不能理解IOU lable怎么会分布到[-1,1]。您说预测得到的iou是预测到的bounding box 与真值框的IOU，怎么会有负数呢？就如完全不匹配顶多是IOU=0呀？

[jbr97]

1. class-aware IoU predictors，类别相关（有类别区分的）的IoU检测器
2. 论文里说到，我们将IoU[0.5, 1]的bbox的IoU值线性变换到[-1,1]，线性变换即F(IoU) = 4*IoU-3

[fengxiuyaun]

谢谢回复哈，但我没找到从什么地方体现出有类别相关呢？

[jbr97]

这应该算是专有名词吧，我翻译的也不准，如果你在搜索引擎里输入class-aware detector or class-agnostic detector，应该能获得非常详尽的解释:)

[vacancy]

On Class-aware IoU predictors.

In short, assuming we have 80 classes in our dataset, the output of the IoU prediction branch is a vector iou of length 80 for each bounding box. After determining its class label c (based on the classification branch), we use iou[c] as the predicted IoU.

BTW, the bounding box regression branch in Faster R-CNN/Mask R-CNN is also "class-aware".

[fengxiuyaun]

oh, thank you very much

[vacancy]

Closing this thread. Feel free to reopen it if you have any further questions.