Updates for utils/fuse_appearance_embeddings_into_shs_dc

[hardikdava]

@yzslab I have fixed few things in this script.

• used detach for saving plys
• updated import with latest code

[hardikdava]

@yzslab Right now, it is loading all the data to cpu/gpu. The script did not work on laptop gpu. I had to use bigger machine to work with large dataset. Can you suggest how can I make it lazy image loading for this script. I suggest as following.

• making lazy dataset loading
• make prunning optional

[yzslab]

@yzslab Right now, it is loading all the data to cpu/gpu. The script did not work on laptop gpu. I had to use bigger machine to work with large dataset. Can you suggest how can I make it lazy image loading for this script. I suggest as following.

• making lazy dataset loading
• make prunning optional

Hi, this script only load cameras into GPU, and this will not consume too much GPU memory. Though pruning is optional, calculating the visibilities is required by weighted fusing, so you must use a machine equipped with a NVIDIA GPU.

To reduce GPU memory consumption:

• Try reducing the --camera-chunk-size if you are using wighted color mode
• Or switch to weighted embedding mode

[yzslab]

@hardikdava Hi, take a look the reveiw comment of internal/utils/gaussian_utils.py above.

[yzslab]

Hi, I have made a modification to your internal/utils/gaussian_utils.py. Check whether it works.

[yzslab]

@hardikdava Hi, take a look the reveiw comment of internal/utils/gaussian_utils.py above.

Sorry, I found I forgot to submit the review yesterday.

[hardikdava]

@yzslab sorry for late reply. Added my comments to your questions.

[yzslab]

Thanks!