DISN: Deep Implicit Surface Network for High-quality Single-view 3D Reconstruction

Please report bugs here and we will publish the bug fix and the latest updates

We just released a rendering datasets with more view variations that contains RGB, albedo, depth and normal 2D images.

Please cite our paper DISN: Deep Implicit Surface Network for High-quality Single-view 3D Reconstruction (NeurIPS 2019)

@incollection{NIPS2019_8340,
title = {DISN: Deep Implicit Surface Network for High-quality Single-view 3D Reconstruction},
author = {Xu, Qiangeng and Wang, Weiyue and Ceylan, Duygu and Mech, Radomir and Neumann, Ulrich},
booktitle = {Advances in Neural Information Processing Systems 32},
editor = {H. Wallach and H. Larochelle and A. Beygelzimer and F. d\textquotesingle Alch\'{e}-Buc and E. Fox and R. Garnett},
pages = {492--502},
year = {2019},
publisher = {Curran Associates, Inc.},
url = {http://papers.nips.cc/paper/8340-disn-deep-implicit-surface-network-for-high-quality-single-view-3d-reconstruction.pdf}
}

Code contact: Qiangeng Xu* and Weiyue Wang*

Installation

    pip install trimesh==2.37.20
    cd {DISN}
    mkdir checkpoint
    cd checkpoint
    wget https://www.dropbox.com/s/2ts7qc9w4opl4w4/SDF_DISN.tar?dl=0
    tar -xvzf SDF_DISN.tar?dl=0
    rm -rf SDF_DISN.tar?dl=0
    cd ..
    mkdir cam_est/checkpoint
    cd cam_est/checkpoint
    wget https://www.dropbox.com/s/hyv4lcvpfu0au9e/cam_DISN.tar?dl=0
    tar -xvzf cam_DISN.tar?dl=0
    rm -rf cam_DISN.tar?dl=0
    cd ../../
    change corresponding libary path in your system in isosurface/LIB_PATH

Demo:

--sdf_res control the resolution of the sampled sdf, default is 64, the larger, the more fine-grained, but slower.

cd {DISN}
source isosurface/LIB_PATH
nohup python -u demo/demo.py --cam_est --log_dir checkpoint/SDF_DISN --cam_log_dir cam_est/checkpoint/cam_DISN --img_feat_twostream --sdf_res 256 &> log/create_sdf.log &

The result is demo/result.obj.

Data Preparation

file location setup:

under preprocessing/info.json, you can change the locations of your data: the neccessary dir for the main model are :

"raw_dirs_v1": {
"mesh_dir": "/ssd1/datasets/ShapeNet/ShapeNetCore.v1/",
"norm_mesh_dir": "/ssd1/datasets/ShapeNet/march_cube_objs_v1/",
"rendered_dir": "/ssd1/datasets/ShapeNet/ShapeNetRendering/",
"renderedh5_dir": "/ssd1/datasets/ShapeNet/ShapeNetRenderingh5_v1/",
"sdf_dir": "/ssd1/datasets/ShapeNet/SDF_v1/"
}

Download ShapeNetCore.v1

download the dataset following the instruction of https://www.shapenet.org/account/ (about 30GB)

cd {your download dir}
wget http://shapenet.cs.stanford.edu/shapenet/obj-zip/ShapeNetCore.v1.zip 
unzip ShapeNetCore.v1.zip -d {your mesh_dir}

Prepare SDF files and marching cube ground truth model

To directly download the generated sdf and model, follow the instruction here To generate sdf files and the reconstructed models by yourself (Please expect the script to run for several hours), please execute the following command lines Our data preparation used this paper Vega: non-linear fem deformable object simulator. Please also cite it if you use our code to generate sdf files
```
mkdir log
cd {DISN}
source isosurface/LIB_PATH
nohup python -u preprocessing/create_point_sdf_grid.py --thread_num {recommend 9} --category {default 'all', but can be single category like 'chair'} &> log/create_sdf.log &

## SDF folder takes about 9.0G, marching cube obj folder takes about 245G
```
Download and generate 2d image h5 files:
- download 2d image following 3DR2N2[https://github.com/chrischoy/3D-R2N2], please cite their paper if you use this image tar file:
```
wget http://cvgl.stanford.edu/data2/ShapeNetRendering.tgz
untar it to {your rendered_dir}
```
- run h5 file generation (about 26 GB) :
```
cd {DISN}
nohup python -u preprocessing/create_img_h5.py &> log/create_imgh5.log &
```

Camera parameters estimation network

train the camera parameters estimation network:

### train the camera poses of the original rendered image dataset. 
nohup python -u cam_est/train_sdf_cam.py --log_dir checkpoint/{your training checkpoint dir} --gpu 0 --loss_mode 3D --learning_rate 2e-5 &> log/cam_3D_all.log &

### train the camera poses of the adding 2 more DoF augmented on the rendered image dataset. 
nohup python -u cam_est/train_sdf_cam.py --log_dir checkpoint/{your training checkpoint dir} --gpu 2 --loss_mode 3D --learning_rate 1e-4 --shift --shift_weight 2 &> log/cam_3D_shift2_all.log &

create h5 file of image and estimated cam parameters:

＃＃＃　Create img_h5 to {renderedh5_dir_est} in your info.json, the default is only generate h5 of test images and cam parameters(about 5.3GB) 
nohup python -u train_sdf_cam.py --img_h5_dir {renderedh5_dir_est} --create --restore_model checkpoint/cam_3D_all --log_dir checkpoint/{your training checkpoint dir} --gpu 0--loss_mode 3D --batch_size 24 &> log/create_cam_mixloss_all.log &

SDF generation network:

train the sdf generation with provided camera parameters:

if train from scratch, you can load official pretrained vgg_16 by setting --restore_modelcnn; or you can --restore_model to your checkpoint to continue the training):
- support flip the background color from black to white since most online images have white background(by using --backcolorwhite)
- if use flag --cam_est, the img_h5 is loaded from {renderedh5_dir_est} instead of {renderedh5_dir}, so that we can train the generation on the estimated camera parameters
```
nohup python -u train/train_sdf.py --gpu 0 --img_feat_twostream --restore_modelcnn ./models/CNN/pretrained_model/vgg_16.ckpt --log_dir checkpoint/{your training checkpoint dir} --category all --num_sample_points 2048 --batch_size 20 --learning_rate 0.0001 --cat_limit 36000 &> log/DISN_train_all.log &
```
inference sdf and create mesh objects:
- will save objs in {your training checkpoint dir}/test_objs/{sdfres+1}{iso}
- will save objs in {your training checkpoint dir}/test_objs/{sdfres+1}{iso}
- if use estimated camera post, --cam_est, will save objs in {your training checkpoint dir}/testobjs/camest{sdfres+1}{iso}
- if only create chair or a single category, --category {chair or a single category}
- --sdf_res control the resolution of the sampled sdf, default is 64, the larger, the more fine-grained, but slower.
```
source isosurface/LIB_PATH
```
use ground truth camera pose

nohup python -u test/create_sdf.py --img_feat_twostream --view_num 24 --sdf_res 64 --batch_size 1 --gpu 0 --sdf_res 64 --log_dir checkpoint/{your training checkpoint dir} --iso 0.00 --category all &> log/DISN_create_all.log &

use estimated camera pose

nohup python -u test/create_sdf.py --img_feat_twostream --view_num 24 --sdf_res 64 --batch_size 1 --gpu 3 --sdf_res 64 --log_dir checkpoint/{your training checkpoint dir} --iso 0.00 --category all --cam_est &> log/DISN_create_all_cam.log &

clean small objects:

if the model doens't converge well, you can clean flying parts that generated by mistakes

nohup python -u clean_smallparts.py --src_dir checkpoint/{your training checkpoint dir}/test_objs/65_0.0 --tar_dir checkpoint/{your training checkpoint dir}/test_objs/65_0.0 --thread_n 10 &> log/DISN_clean.log &

Evaluation:

please compile models/tf_ops/ approxmatch and nn_distance and cites "A Point Set Generation Network for 3D Object Reconstruction from a Single Image"

Chamfer Distance and Earth Mover Distance:

cal_dir specify which obj folder to be tested, e.g. if only test watercraft, --category watercraft

nohup python -u test/test_cd_emd.py --img_feat_twostream --view_num 24 --num_sample_points 2048 --gpu 0 --batch_size 24 --log_dir checkpoint/{your training checkpoint dir} --cal_dir checkpoint/{your training checkpoint dir}/test_objs/65_0.0 --category all &> log/DISN_cd_emd_all.log &

F-Score caluculation:

cal_dir specify which obj folder to be tested, e.g. if only test watercraft, --category watercraft also the threshold of true can be set, here we use 2.5 for default:

nohup python -u test/test_f_score.py --img_feat_twostream --view_num 24 --num_sample_points 2048 --gpu 0 --batch_size 24 --log_dir checkpoint/{your training checkpoint dir} --cal_dir checkpoint/{your training checkpoint dir}/test_objs/65_0.0 --category all --truethreshold 2.5 &> log/DISN_fscore_2.5.log &

IOU caluculation:
cal_dir specify which obj folder to be tested, e.g. if only test watercraft, --category watercraft
--dim specify the number of voxels along each 3D dimension.

  nohup python -u test/test_iou.py --img_feat_twostream --view_num 24 --log_dir checkpoint/{your training checkpoint dir} --cal_dir checkpoint/{your training checkpoint dir}/test_objs/65_0.0 --category all --dim 110 &> DISN_iou_all.log &

laughtervv / DISN

readme

DISN: Deep Implicit Surface Network for High-quality Single-view 3D Reconstruction

We just released a rendering datasets with more view variations that contains RGB, albedo, depth and normal 2D images.

Installation

Demo:

Data Preparation

file location setup:

Download ShapeNetCore.v1

Prepare SDF files and marching cube ground truth model

Download and generate 2d image h5 files:

download 2d image following 3DR2N2[https://github.com/chrischoy/3D-R2N2], please cite their paper if you use this image tar file:

run h5 file generation (about 26 GB) :

Camera parameters estimation network

train the camera parameters estimation network:

create h5 file of image and estimated cam parameters:

SDF generation network:

train the sdf generation with provided camera parameters:

inference sdf and create mesh objects:

use ground truth camera pose

use estimated camera pose