LiuYi-Up

任务申领：二、优化 ncnn int8 模型部署 微信号：ly1150854440

【队名】：放平心态 【序号】：10 【论文】：[Per-Pixel Classification is Not All You Need for Semantic Segmentation](https://arxiv.org/abs/2107.06278) 【状态】：报名 【repo链接】：https://github.com/LiuYi-Up/PdPaper-MaskFormer

奥！非常感谢！   ------------------ 原始邮件 ------------------ 发件人: "Elody-07/AWR-Adaptive-Weighting-Regression" ***@***.***>; 发送时间: 2022年4月5日(星期二) 上午10:16 ***@***.***>; 抄送: "惯 ***@***.******@***.***>; 主题: Re: [Elody-07/AWR-Adaptive-Weighting-Regression] 关于eval_tool.feed()逆变换的问题 (Issue #18) 因为网络输出的结果是图像坐标系下的，并进行了一系列数据增强变换，而ground truth是相机坐标系下的，具体可以参考论文Spatial-aware stacked regression network for real-time 3D hand pose estimation中Coordinate Decoupling的介绍 — Reply...

Hi Bardia Doosti!I trained it for about 2000 epoch on HO3D, but the 2D results did not seem to be good. They were generally shrunk and were close to the...

> Hi, Liu! Sorry, I can't share your questions meaningfully at present, but I don't know if I can ask you some questions, because I'm a green, so there are...

My configuration environment: python 3.7.10 torch 1.2.0 cuda V10.0.130

> @LiuYi-Up 您好，请问您的问题解决了吗？我也遇到同样的问题，使用公开数据集训练集上效果很差，预测点大多都集中于图像中间。这是数据输入的问题吗？还是网络结构的问题呢？ 您好，我并没有解决这个问题，我的猜想是网络前需要手的检测器crop后再输入HOPE-Net，但由于各种原因，我并没有继续进行实验。

> Hi, I found that the official FPHA dataset only provide joints 3d location for hand pose annotations. I wonder how did you train on HandOccNet? I use the points...

> > > Hi, I found that the official FPHA dataset only provide joints 3d location for hand pose annotations. I wonder how did you train on HandOccNet? > >...