HibikiJie / LicensePlate

License plate recognition
62 stars 7 forks source link

您好,代码中用到的resnet18网络,我没有在原论文的介绍中看到啊,能不能麻烦你详细介绍一下wpodnet那部分 #5

Open bqtd opened 4 months ago

HibikiJie commented 4 months ago

resnet18作为特征提取器,生成的特征图上,以YOLO的检测原理,每一个cell中,有目标object的置信度,以及相应预测的仿射变换矩阵。

bqtd commented 4 months ago

请问答主这边有训练好的表现比较好的权重文件嘛?原代码的这个感觉表现比较差

---Original--- From: @.> Sent at: 2024年5月7日(Tue) AM10:56 To: @.>; Cc: @.**@.>; Subject: Re: [HibikiJie/LicensePlate] 您好,代码中用到的resnet18网络,我没有在原论文的介绍中看到啊,能不能麻烦你详细介绍一下wpodnet那部分 (Issue #5)

resnet18作为特征提取器,生成的特征图上,以YOLO的检测原理,每一个cell中,有目标object的置信度,以及相应预测的仿射变换矩阵。

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

HibikiJie commented 3 months ago

可以自行添加真实车牌数据训练

bqtd commented 3 months ago

我自己训练的效果并不是那么好,由于水平有限,所以很难找出问题在哪

---Original--- From: @.> Sent at: 2024年5月22日(Wed) AM10:18 To: @.>; Cc: @.**@.>; Subject: Re: [HibikiJie/LicensePlate] 您好,代码中用到的resnet18网络,我没有在原论文的介绍中看到啊,能不能麻烦你详细介绍一下wpodnet那部分 (Issue #5)

可以自行添加真实车牌数据训练

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

HibikiJie commented 3 months ago

CCPD2019的数据几乎只在一个地方采集,所以车牌检车会较弱。而且车牌号的识别全部使用合成数据产生,对于unseen的数据可能适应性不好。这需要额外的真实数据。

bqtd commented 3 months ago

好的,感谢解答

---Original--- From: @.> Sent at: 2024年5月22日(Wed) AM10:26 To: @.>; Cc: @.**@.>; Subject: Re: [HibikiJie/LicensePlate] 您好,代码中用到的resnet18网络,我没有在原论文的介绍中看到啊,能不能麻烦你详细介绍一下wpodnet那部分 (Issue #5)

CCPD2019的数据几乎只在一个地方采集,所以车牌检车会较弱。而且车牌号的识别全部使用合成数据产生,对于unseen的数据可能适应性不好。这需要额外的真实数据。

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>