Skip to content

Latest commit

 

History

History
254 lines (187 loc) · 18.1 KB

README.md

File metadata and controls

254 lines (187 loc) · 18.1 KB
Raw

📢 NEWS: We have released MVEdit, an upgraded codebase based on SSDNeRF. MVEdit supports all SSDNeRF models and configs, and offers new features such as diffusers support and improved SSDNeRF GUI.

SSDNeRF

Official PyTorch implementation of the ICCV 2023 paper:

Single-Stage Diffusion NeRF: A Unified Approach to 3D Generation and Reconstruction
Hansheng Chen1,*, Jiatao Gu2, Anpei Chen3, Wei Tian1, Zhuowen Tu4, Lingjie Liu5, Hao Su4
1Tongji University, 2Apple, 3ETH Zürich, 4UCSD, 5University of Pennsylvania
*Work done during a remote internship with UCSD.

[project page] [paper]

Part of this codebase is based on torch-ngp and MMGeneration.

teaser.mp4

Highlights

  • Code to reproduce ALL the experiments in the paper and supplementary material (including single-view reconstruction on the real KITTI Cars dataset).
  • New features including support for tiled triplanes (rollout layout), FP16 diffusion sampling, and 16-bit caching.
  • A simple GUI demo (modified from torch-ngp).

Installation

Prerequisites

The code has been tested in the environment described as follows:

Also, this codebase should be able to work on Windows systems as well (tested in the inference mode).

Other dependencies can be installed via pip install -r requirements.txt.

An example of commands for installing the Python packages is shown below:

# Export the PATH of CUDA toolkit
export PATH=/usr/local/cuda-11.3/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda-11.3/lib64:$LD_LIBRARY_PATH

# Create conda environment
conda create -y -n ssdnerf python=3.7
conda activate ssdnerf

# Install PyTorch (this script is for CUDA 11.3)
conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch

# Install MMCV and MMGeneration
pip install -U openmim
mim install mmcv-full==1.6
git clone https://github.com/open-mmlab/mmgeneration && cd mmgeneration && git checkout v0.7.2
pip install -v -e .
cd ..

# Clone this repo and install other dependencies
git clone <this repo> && cd <repo folder>
pip install -r requirements.txt

Compile CUDA packages

There are two CUDA packages from torch-ngp that need to be built locally.

cd lib/ops/raymarching/
pip install -e .
cd ../shencoder/
pip install -e .
cd ../../..

Data preparation

Download srn_cars.zip and srn_chairs.zip from here. Unzip them to ./data/shapenet.

Download abo_tables.zip from here. Unzip it to ./data/abo. For convenience I have converted the ABO dataset into PixelNeRF's SRN format.

If you want to try single-view reconstruction on the real KITTI Cars dataset, please download the official KITTI 3D object dataset, including left color images, calibration files, training labels, and instance segmentations.

Extract the downloaded archives according to the following folder tree (or use symlinks).

./
├── configs/
├── data/
│   ├── shapenet/
│   │   ├── cars_test/
│   │   ├── cars_train/
│   │   ├── cars_val/
│   │   ├── chairs_test/
│   │   ├── chairs_train/
│   │   └── chairs_val/
│   ├── abo/
│   │   ├── tables_train/
│   │   └── tables_test/
│   └── kitti/
│       └── training/
│           ├── calib/
│           ├── image_2/
│           ├── label_2/
|           └── instance_2/
├── demo/
├── lib/
├── tools/
…

For FID and KID evaluation, run the following commands to extract the Inception features of the real images. (This script will use all the available GPUs on your machine, so remember to set CUDA_VISIBLE_DEVICES.)

CUDA_VISIBLE_DEVICES=0 python tools/inception_stat.py configs/paper_cfgs/ssdnerf_cars_uncond.py
CUDA_VISIBLE_DEVICES=0 python tools/inception_stat.py configs/paper_cfgs/ssdnerf_chairs_recons1v.py
CUDA_VISIBLE_DEVICES=0 python tools/inception_stat.py configs/paper_cfgs/ssdnerf_abotables_uncond.py

For KITTI Cars preprocessing, run the following command.

python tools/kitti_preproc.py

About the configs

Naming convention

ssdnerf_cars3v_uncond
   │      │      └── testing data: test unconditional generation
   │      └── training data: train on Cars dataset, using 3 views per scene
   └── training method: single-stage diffusion nerf training
  
stage2_cars_recons1v
   │     │      └── testing data: test 3D reconstruction from 1 view
   │     └── training data: train on Cars dataset, using all views per scene
   └── training method: stage 2 of two-stage training

Models in the main paper

Config Checkpoint Iters FID LPIPS Comments
ssdnerf_cars_uncond gdrive 1M 11.08 ± 1.11 -
ssdnerf_abotables_uncond gdrive 1M 14.27 ± 0.66 -
ssdnerf_cars_recons1v gdrive 80K 16.39 0.078
ssdnerf_chairs_recons1v gdrive 80K 10.13 0.067
ssdnerf_cars3v_uncond_1m 1M - The first half of training before resetting the triplanes.
ssdnerf_cars3v_uncond_2m gdrive 1M 19.04 ± 1.10 - The second half of training after resetting the triplanes (requires training ssdnerf_cars3v_uncond_1m first).
ssdnerf_cars3v_recons1v 80K 0.106
stage1_cars_recons16v 400K Ablation study, NeRF reconstruction stage.
stage2_cars_uncond 1M 16.33 ± 0.93 - Ablation study, diffusion stage (requires training stage1_cars_recons16v first).
stage2_cars_recons1v 80K 20.97 0.090 Ablation study, diffusion stage (requires training stage1_cars_recons16v first).

In addition, multi-view reconstruction testing configs can be found in configs/paper_cfgs/multiview_recons.

Models in the supplementary material

Config Iters FID LPIPS Comments
ssdnerf_cars_reconskitti 80K - - Same model as ssdnerf_cars_recons1v [checkpoint] except for being tested on real images of the KITTI dataset.
ssdnerf_cars_recons1v_notanh 80K 16.34 0.077 Without tanh latent code activation.
ssdnerf_cars_recons1v_noreg 80K 16.62 0.077 Without L2 latent code regularization.

New models in this repository

The new models feature improved implementations, including the following changes:

  • Use NormalizedTanhCode instead of TanhCode activation, which helps stablizing the scale (std) of the latent codes. Scale normalization is no longer required in the DDPM MSE loss. Latent code lr is rescaled accordingly.
  • Remove L2 latent code regularizaiton.
  • Disable U-Net dropout in recons models.
  • uncond and recons models are now exactly the same except for training schedules and testing configs.
  • Enable new features such as 16-bit caching and tiled triplanes.

Note: It is highly recommended to start with these new models if you want to train custom models. The original models in the paper are retained only for reproducibility.

Config Iters Comments
ssdnerf_cars_uncond_16bit 1M Enable 16-bit caching. Should yield similar results to ssdnerf_cars_uncond.
ssdnerf_cars_recons1v_16bit 60K Enable 16-bit caching. Should yield similar results to ssdnerf_cars_recons1v.
ssdnerf_cars_recons1v_tiled 100K Use tiled (rollout) triplane layout. Tiled triplanes could have resulted in higher computation cost, but in this model the UNet channels have been reduced to compensate for the runtime.
stage1_cars_recons16v_16bit 400K Enable 16-bit caching. Should yield similar results to stage1_cars_recons16v.
stage1_cars_recons16v_16bit_filesystem 400K Same as stage1_cars_recons16v_16bit but caching on filesystem, in case your RAM is full. Not recommended due to slow I/O on hard drives.

Unused features in this codebase

  • This codebase supports concat-based image conditioning, although it's not used in the above models.

Training

Run the following command to train a model:

python train.py /PATH/TO/CONFIG --gpu-ids 0 1

Note that the total batch size is determined by the number of GPUs you specified. All our models are trained using 2 RTX 3090 (24G) GPUs.

Since we adopt the density-based NeRF pruning trategy in torch-ngp, training would start slow and become faster later, so the initial esitamtion of remaining time is usually over twice as much as the actual training time.

Model checkpoints will be saved into ./work_dirs. Scene caches will be saved into ./cache.

Testing and evaluation

python test.py /PATH/TO/CONFIG /PATH/TO/CHECKPOINT --gpu-ids 0 1  # you can specify any number of GPUs here

Some trained models can be downloaded from here for testing.

To save the sampled NeRFs and extracted meshes, uncomment (or add) these lines in the test_cfg dict of the config file:

    save_dir=work_dir + '/save',
    save_mesh=True,
    mesh_resolution=256,
    mesh_threshold=10,

All results will be saved into ./work_dirs/<cfg name>/save. You can then open the saved .pth NeRF scenes using the GUI tool demo/ssdnerf_gui.py (see below), and the .stl meshes using any mesh viewer.

Visualization

By default, during training or testing, the visualizations will be saved into ./work_dirs.

A GUI tool is provided for visualizing the results (currently only supports unconditional generation or loading saved .pth NeRF scenes). Run the following command to start the GUI:

python demo/ssdnerf_gui.py /PATH/TO/CONFIG /PATH/TO/CHECKPOINT --fp16

Citation

If you find this project useful in your research, please consider citing:

@inproceedings{ssdnerf,
    title={Single-Stage Diffusion NeRF: A Unified Approach to 3D Generation and Reconstruction}, 
    author={Hansheng Chen and Jiatao Gu and Anpei Chen and Wei Tian and Zhuowen Tu and Lingjie Liu and Hao Su},
    year={2023},
    booktitle={ICCV}
}