congjianting

@abc8350712 i face the same thing, did you find why?

@blacklong617 did you fintune the model using vgg16.ckpt with pascalvoc 2007 ?

@blacklong617 my result is very bad, the loss cannot decrease, if you do not finetune vgg_16.ckpt, how do you use this project for other dataset?

@blacklong617 the class number changed, meaning that you need to exclude some layers to finetune.

Hi, the following link is not available. Please fix it, thanks. You can download all the models from our [model zoo](https://1drv.ms/u/s!AhIXJn_J-blW9RzF3rMW7SsLHa8h?e=blQ0Al).

哦, 希望老师们考虑考虑, 感觉有些相邻图片即使中心位置的位移不大, 但目标尺度变化剧烈时, 回归的bbox也是不太准确的, 如果能像cascade的回归方式让bbox越来越准就更好了.

个人感受这类现象的图片序列即使在训练上难度也比较大, 收敛的bbox和score都不是很好的, 我是基于cvt的模型做的.

@yutaocui 调整searchareascale参数确实能缓解下 , 但同时带来了尺度,位移相对平缓变化的图像序列的回归精度下降, 表现在search的图的推理目标在score上挺高的, 但bbox的边界不太准确, 因此实际工作中也挺难平衡这个矛盾的.

hi,  i want to train a model. My friends have downloaded the cvt-21 pretrained model from the model zoo link. thanks. ------------------ 原始邮件 ------------------ 发件人: "aehrc/cvt2distilgpt2" ***@***.***>; 发送时间: 2022年11月7日(星期一) 中午12:59 ***@***.***>; ***@***.******@***.***>; 主题: Re:...