search

MVIG-SJTU / pointSIFT

a module for 3D semantic segmentation in point clouds.
644 stars 119 forks source link

How is the effect of training? #17

Open overfitover opened 6 years ago

overfitover commented 6 years ago

Could you provide a trained model?

overfitover commented 6 years ago

@zhijian-liu @YuliangXiu @Vera-He @songzhaozhe Do you get the result?

meteorshowers commented 6 years ago

I get a bad result on scannet about 84.7 use your code strictly, but your paper's report is 86. could you provide the trained model and explain how to get the result use your code?

overfitover commented 6 years ago

@meteorshowers 你的结果怎么样？不是这种数字，是预测的可视化效果怎么样？我觉得这数字上的差异毕竟不是太大可以接受的。毕竟我们只是局部最优解，他可能是训练了很多次找到了更好的局部最优解。你也可以多训练几次，也许会有提升。

meteorshowers commented 6 years ago

可以@下作者

LaviLiu commented 6 years ago

@meteorshowers 训练的时候只需要像素的坐标点就可以了吗，不需要像素值吗？我看到代码里好像没有用到像素值

overfitover commented 6 years ago

@Laviyy 点云的就是点的坐标位置。你理解的像素值是什么？他其实是提取每个点和周围点的特征，让其具有局部特征，以弥补pointnet的不足。

928776539 commented 3 years ago

Is there code implemented by Pytorch？