关于3D视觉的注意事项

最新推荐文章于 2024-12-02 16:19:15 发布

_Gus_

最新推荐文章于 2024-12-02 16:19:15 发布

阅读量193

点赞数

文章标签： 3d

本文链接：https://blog.youkuaiyun.com/Gussss/article/details/116594237

版权

假如相机内参不确定，可以简单设置为fx = cx = w/2 and fy = cy = h / 2. https://github.com/TRI-ML/packnet-sfm/issues/38

估计内参论文 https://arxiv.org/abs/2008.06630

点云可视化https://github.com/TRI-ML/packnet-sfm/issues/10

city预训练，只要改内参https://github.com/nianticlabs/monodepth2/issues/319

在这里插入图片描述

https://blog.youkuaiyun.com/try_again_later/article/details/10545 6252

https://github.com/nianticlabs/monodepth2/issues/238

Hi,

We have batchnorm in the encoder because resnet18 has it, however as shown in the groupnorm paper, training with both small batch sizes (< 32 which is the case for monodepth2) and batchnorm hurts accuracy.
However I don't recall adding batchnorm and comparing, it's actually possible that it could improve results.

I hope this helps!

点云，mesh 和体素体素和点云的区别是什么？ - 知乎