Pytorch3d: Quantitative performance measure for the NERF implementation

[zhangmozhe]

❓ Quantitative performance measure for the NERF implementation

Thanks for releasing the NERF implementation using Open3D. I wonder whether the quantitative performance matches the original implementation? Adding such quantitative measurement will definitely help increase the impact for the project. And It will be nice to add a description about whether there is any simplification or difference in this implementation.

BTW, the S3 bucket link for the dataset file (https://dl.fbaipublicfiles.com/pytorch3d_nerf_data) does not work.

Thanks for the explanation!

[nikhilaravi]

@zhangmozhe you can access the data files following this format:

https://github.com/facebookresearch/pytorch3d/blob/ff9c6612b457a2021d88fea119bdb9b94ba017bd/projects/nerf/nerf/dataset.py#L148-L155

e.g. the lego cameras are at: https://dl.fbaipublicfiles.com/pytorch3d_nerf_data/lego.pth.

@davnov134 compared with the original NeRF implementation and can provide more details about performance.

[zhangmozhe]

Cool. Thanks for your reply.

[zhangmozhe]

When I open the link (https://dl.fbaipublicfiles.com/pytorch3d_nerf_data/lego_cameras.pth), it still prompts:

[bottler]

It's https://dl.fbaipublicfiles.com/pytorch3d_nerf_data/lego.pth . The code for downloading it is right.

[nikhilaravi]

@bottler thanks, I have updated my previous comment with the correction!

[davnov134]

❓ Quantitative performance measure for the NERF implementation

Thanks for releasing the NERF implementation using Open3D. I wonder whether the quantitative performance matches the original implementation? Adding such quantitative measurement will definitely help increase the impact for the project. And It will be nice to add a description about whether there is any simplification or difference in this implementation.

BTW, the S3 bucket link for the dataset file (https://dl.fbaipublicfiles.com/pytorch3d_nerf_data) does not work.

Thanks for the explanation!

@zhangmozhe we have now updated the readme file with a quantitative comparison to the official TensorFlow implementation.

[zhangmozhe]

Great! Thanks for your update about this! BTW, is there any plan to support multi-GPU training using DDP?

[davnov134]

It is not on our immediate roadmap, but we will definitely consider this as it seems like a very useful addition.

[nikhilaravi]

@zhangmozhe thanks for the suggestion. We will definitely plan to add this support. I will close this issue now that the performance information has been added.