NU-MCC: Multiview Compressive Coding with Neighborhood Decoder and Repulsive UDF

This repository contains the official implementation of the paper:

NU-MCC: Multiview Compressive Coding with Neighborhood Decoder and Repulsive UDF
Stefan Lionar, Xiangyu Xu, Lin Min, and Gim Hee Lee
Accepted at NeurIPS 2023

Installation

Please see INSTALL.md for information on installation.

Data

Please see DATASET.md for information on data preparation.

Pretrained models

To download the pretrained models, run:

mkdir pretrained

# CO3D-V2 Repulsive UDF
wget https://numcc.s3.us-west-1.amazonaws.com/udf-ep99.pth -P pretrained

# Hypersim
wget https://numcc.s3.us-west-1.amazonaws.com/numcc_hypersim_550c.pth -P pretrained

Zero-shot Demo

We provide demo for reconstruction from iPhone capture and AI-generated image.

To run demo for reconstruction from iPhone capture, run:

python demo_iphone.py

For AI-generated image, run:

python demo_web.py

Output visualization will be generated to demo/output.html.

CO3D-v2 Experiments

To train NU-MCC from scratch, run:

PYTHONHASHSEED=[SEED] torchrun --nproc_per_node [NUM_GPU] main_numcc.py --exp_name [YOUR_EXPERIMENT_NAME] --accum_iter [32/NUM_GPU]

For example, to train with 4 GPUs:

PYTHONHASHSEED=0 torchrun --nproc_per_node 4 main_numcc.py --exp_name numcc_udf --accum_iter 8

For evaluation/inference:

# Standard inference
PYTHONHASHSEED=[SEED] torchrun --nproc_per_node [N_GPU] main_numcc.py --run_val --resume [MODEL_PATH] --n_query_udf [BATCH_QUERY_FOR_REPULSIVE]

# High-resolution
PYTHONHASHSEED=[SEED] torchrun --nproc_per_node [N_GPU] main_numcc.py --run_val --resume [MODEL_PATH] --n_query_udf [BATCH_QUERY_FOR_REPULSIVE] --hr --xyz_size_hr 224

# Smoothing
PYTHONHASHSEED=[SEED] torchrun --nproc_per_node [N_GPU] main_numcc.py --run_val --resume [MODEL_PATH] --n_query_udf [BATCH_QUERY_FOR_REPULSIVE] --nneigh 12 --nn_seen 12

PYTHONHASHSEED defines the random seed for seen images. The argument --n_query_udf defines the batch query points for the repulsive force. In general, the higher numbers result in more uniform point distribution, but somewhere around 10k points is already good enough.

An example of a working command with 4 GPUs:

PYTHONHASHSEED=0 torchrun --nproc_per_node 4 main_numcc.py --run_val --resume pretrained/udf-ep99.pth --n_query_udf 48000

To run visualization, use --run_viz flag. The output will be generated to the folder specified in --exp_name. Visualization/evaluation from one class can be specified using --one_class [OBJECT_CLASS] flag. Point clouds can be exported by activating --save_pc flag.

Hypersim Experiment

To train on Hypersim dataset, run:

torchrun --nproc_per_node 4 main_numcc.py --exp_name [EXPERIMENT_NAME] --hypersim_path [DATASET_PATH] --use_hypersim --blr 5e-5 --epochs 50 --train_epoch_len_multiplier 3200 --accum_iter 8 --n_groups 550

For evaluation/visualization, use --run_val or --run_viz flags and specify the model on --resume [MODEL_PATH].

Note: Outputs from the data preparation, i.e., hypersim_gt_train.pt and hypersim_gt_val.pt need to be placed in this repository's home directory.

Acknowledgement

This codebase is mainly inherited from Meta Platforms' MCC codebase.

Citation

If you find our code or paper useful, please consider citing us:

@inproceedings{lionar2023nu,
  title={NU-MCC: Multiview Compressive Coding with Neighborhood Decoder and Repulsive UDF},
  author={Lionar, Stefan and Xu, Xiangyu and Lin, Min and Lee, Gim Hee},
  booktitle={Advances in neural information processing systems},
  year={2023}
}

sail-sg / numcc