xuechaoyan comments

Results 6 comments of


                                            xuechaoyan

Question: Reducing False Positives with Empty annotation Images in training process

亲爱的作者大大，我在尝试使用空标签的图像来训练的时候会在epoch1卡住，但是也不报错，是不是我们的算法不能使用空标签的图片来训练？QAQ求指导一下。

Question: Reducing False Positives with Empty annotation Images in training process

> Yes, this approach works because the loss will be computed on false activations, which encourages the model to make fewer mistakes, assuming the data is homoscedastic and from similar...

Question: Reducing False Positives with Empty annotation Images in training process

> 添加空白图片参与训练和denoising.py在与图片原有目标没有交集的区域随机生成目标，是可以缓解误报，但对部分数据集还是没有很好的解决误报问题。如果在dec_socre_head中与detr类似类别数量增加一类当作背景，把随机生成的框标签视为背景类，在分类时会不会把之前与待检测目标不相似的误报归类于背景，减少误报的产生，对检出率影响有多大我在我的实验中是并没有设定多一个背景的类，这个误报缓解具体看你的误报图片，你可以把你的结果生成出来看，看看算法具体在哪些地方，或者在什么图片情况下容易误报，然后把这类的图片当成没有标签的图片拿去训练。在我的实验场景下，这样可以降低很多的误报

转换成coco数据集格式的时候，标签起始id是0还是1？

同问啊，我也是这个问题

COCO category id has to start wiht 0 but should start with 1.

Does it work when you changed the coco category ids with 0? How's the weight you trained?

Too high false positives

> Hey have you all tried reducing the `num_queries` parameter? The default is 300 and if you are training on a custom dataset that has less objects per image, this...