Results 10 comments of Xue Wen

I have the same question. Why not 64X48 [32 channels] -> 32X24 [64 channels] -> 16X12 [128 channels]?

According to my understanding, In "Deep High-Resolution Representation Learning for Human Pose Estimation", a "exchange block" is equal "modularized block" in "Deep High-Resolution Representation Learning for Visual Recognition". The modularized...

@xiaofengShi 大神,您好!我在跑demo的是否遇到了下面的错误,能否麻烦您知道一下呢?谢谢您! Using TensorFlow backend. Tensor("Placeholder:0", shape=(?, ?, ?, 3), dtype=float32) Tensor("conv5_3/conv5_3:0", shape=(?, ?, ?, 512), dtype=float32) Tensor("rpn_conv/3x3/rpn_conv/3x3:0", shape=(?, ?, ?, 512), dtype=float32) Tensor("lstm_o/Reshape_2:0", shape=(?, ?, ?, 512), dtype=float32) Tensor("lstm_o/Reshape_2:0",...

您好!可以将Adam置为True试试。我写一下我的调参路径,希望对您有用。 1、参数都没改,用自己的数据集,训练100个epoch,最后的P、R一直都不高; ![results](https://user-images.githubusercontent.com/8676098/200528734-6017f437-4fe7-4e8d-834b-b5c56f765fa3.png) 2、后来我参考 https://github.com/hukaixuan19970627/yolov5_obb/issues/380 ,调小了lr0: 0.001,mosaic: 0.0,使用exp2的best作为预训练模型;训练到100早停了; ![results](https://user-images.githubusercontent.com/8676098/200530189-51a8c907-ebda-4dbe-aec6-7c617955a174.png) 3、参考 https://github.com/hukaixuan19970627/yolov5_obb/issues/57#issuecomment-888931205 ,Adam置为True,使用exp2的best作为预训练模型,lr0: 0.001,mosaic: 0.0,P、R提升很大 ![results](https://user-images.githubusercontent.com/8676098/200530501-1b61ca9a-d4e6-4ddb-ac43-41b82cd04943.png)

@lyx599 我认为您的思路是对的。谢谢!

> 您好,打扰您了。我想问一下,这个项目更新之后使用的是DOTA数据集格式进行训练。那么现在旋转物体标注方式是不是就是八参数法了(还是说是DOTA格式的任意四边形法),后续的旋转检测优化是不是就是基于八参数法进行的了。还是说使用DOTA格式进行训练和使用YOLO格式训练一样,都是基于五参数法(长边表示法)进行旋转检测优化的呀 您好!这个项目准备的label格式是8参数,加载训练数据前会将利用最小外切矩形将4点坐标坐标转化成五参数法(长边表示法),然后用于训练。

> 请问下,作者似乎对排坑后的DOTA格式转YOLO格式代码修改并上传了,是把脚本文件名改为poly2rbox了吗?直接用opencv里的最小外切矩形是不是还会有作者说的问题。 您说得是那个代码呢? 我提issue的部分,是将数据集中labelTxt中的旋转框4点坐标转换成角度,用于模型学习的那一块。

有可能是没有加激活函数的缘故吧? ![image](https://github.com/user-attachments/assets/d01b23cc-853f-4f81-89e4-af3a793d41af)

> hi--- thank you for your interest. > > vectors u and v are a flattened, absolute-valued, and sorted version of the weight matrix at each layer. (i wrote only...

> This is not an error. Both the original expression and the one that you suggested should work, technically. > > In the above thread, I used descending order because...