Holistic 3D Scene Understanding from a Single Image with Implicit Representation

Implicit3DUnderstanding

Holistic 3D Scene Understanding from a Single Image with Implicit Representation
Cheng Zhang*, Zhaopeng Cui*, Yinda Zhang*, Shuaicheng Liu, Bing Zeng, Marc Pollefeys

img--2-

3dbbox--2-

recon--2-

img--1-

3dbbox--1-

recon--1-

3dbbox

recon

pipeline

Introduction

This repo contains training, testing, evaluation, visualization code of our CVPR 2021 paper.
Specially, the repo contains our PyTorch implementation of the decoder of LDIF, which can be extracted and used in other projects.
We are expecting to release a refactored version of our pipeline and a PyTorch implementation of the full LDIF model in the future.

Install

Please make sure to install CUDA NVCC on your system first. then run the following:

sudo apt install xvfb ninja-build freeglut3-dev libglew-dev meshlab
conda env create -f environment.yml
conda activate Im3D
python project.py build

When running python project.py build, the script will run external/build_gaps.sh which requires password for sudo privilege for apt-get install.
Please make sure you are running with a user with sudo privilege.
If not, please reach your administrator for installation of these libraries and comment out the corresponding lines then run python project.py build.

Demo

Download the pretrained checkpoint
and unzip it into out/total3d/20110611514267/
Change current directory to Implicit3DUnderstanding/ and run the demo, which will generate 3D detection result and rendered scene mesh to demo/output/1/
```
CUDA_VISIBLE_DEVICES=0 python main.py out/total3d/20110611514267/out_config.yaml --mode demo --demo_path demo/inputs/1
```

In case you want to run it off screen (for example, with SSH)

CUDA_VISIBLE_DEVICES=0 xvfb-run -a -s "-screen 0 800x600x24" python main.py out/total3d/20110611514267/out_config.yaml --mode demo --demo_path demo/inputs/1

If you want to run it interactively, change the last line of demo.py

scene_box.draw3D(if_save=True, save_path = '%s/recon.png' % (save_path))

scene_box.draw3D(if_save=False, save_path = '%s/recon.png' % (save_path))

Data preparation

We follow Total3DUnderstanding to use SUN-RGBD to train our Scene Graph Convolutional Network (SGCN), and use Pix3D to train our Local Implicit Embedding Network
(LIEN) with Local Deep Implicit Functions (LDIF) decoder.

Preprocess SUN-RGBD data

Please follow Total3DUnderstanding to directly download the processed train/test data.

In case you prefer processing by yourself or want to evaluate 3D detection with our code
(To ultilize the evaluation code of Coop, we modified the data processing code of Total3DUnderstanding to save parameters for transforming the coordinate system from Total3D back to Coop),
please follow these steps:

Follow Total3DUnderstanding to download the raw data.
According to issue #6 of Total3DUnderstanding,
there are a few typos in json files of SUNRGBD dataset, which is mostly solved by the json loader.
However, one typo still needs to be fixed by hand.
Please find {"name":""propulsion"tool"} in data/sunrgbd/Dataset/SUNRGBD/kv2/kinect2data/002922_2014-06-26_15-43-16_094959634447_rgbf000089-resize/annotation2Dfinal/index.json and remove ""propulsion.
Process the data by
```
python -m utils.generate_data
```

Preprocess Pix3D data

We use a different data process pipeline with Total3DUnderstanding. Please follow these steps to generate the train/test data:

Download the Pix3D dataset to data/pix3d/metadata
Run below to generate the train/test data into 'data/pix3d/ldif'
```
python utils/preprocess_pix3d4ldif.py
```

Training and Testing

We use wandb for logging and visualization.
You can register a wandb account and login before training by wandb login.
In case you don't need to visualize the training process, you can put WANDB_MODE=dryrun before the commands bellow.

Thanks to the well-structured code of Total3DUnderstanding, we use the same method to manage parameters of each experiment with configuration files (configs/****.yaml).
We first follow Total3DUnderstanding to pretrain each individual module, then jointly finetune the full model with additional physical violation loss.

Pretraining

We use the pretrained checkpoint of Total3DUnderstanding to load weights for ODN.
Please download and rename the checkpoint to out/pretrained_models/total3d/model_best.pth.
Other modules can be trained then tested with the following steps:

Train LEN by:
```
python main.py configs/layout_estimation.yaml
```
The pretrained checkpoint can be found at out/layout_estimation/[start_time]/model_best.pth
Train LIEN + LDIF by:
```
python main.py configs/ldif.yaml
```
The pretrained checkpoint can be found at out/ldif/[start_time]/model_best.pth
(alternatively, you can download the pretrained model here, and unzip it into out/ldif/20101613380518/)

The training process is followed with a quick test without ICP and Chamfer distance evaluated. In case you want to align mesh and evaluate the Chamfer distance during testing:
```
python main.py configs/ldif.yaml --mode train
```
The generated object meshes can be found at out/ldif/[start_time]/visualization
Replace the checkpoint directories of LEN and LIEN in configs/total3d_ldif_gcnn.yaml with the checkpoints trained above, then train SGCN by:
```
python main.py configs/total3d_ldif_gcnn.yaml
```
The pretrained checkpoint can be found at out/total3d/[start_time]/model_best.pth

Joint finetune

Replace the checkpoint directory in configs/total3d_ldif_gcnn_joint.yaml with the one trained in the last step above, then train the full model by:
```
python main.py configs/total3d_ldif_gcnn_joint.yaml
```
The trained model can be found at out/total3d/[start_time]/model_best.pth
The training process is followed with a quick test without scene mesh generated. In case you want to generate the scene mesh during testing (which will cost a day on 1080ti due to the unoptimized interface of LDIF CUDA kernel):
```
python main.py configs/total3d_ldif_gcnn_joint.yaml --mode train
```
The testing resaults can be found at out/total3d/[start_time]/visualization

Testing

The training process above already include a testing process. In case you want to test LIEN+LDIF or full model by yourself:
```
python main.py out/[ldif/total3d]/[start_time]/out_config.yaml --mode test
```
The results will be saved to out/total3d/[start_time]/visualization and the evaluation metrics will be logged to wandb as run summary.

Evaluate 3D object detection with our modified matlab script from Coop:

external/cooperative_scene_parsing/evaluation/detections/script_eval_detection.m

Before running the script, please specify the following parameters:

SUNRGBD_path = 'path/to/SUNRGBD';
result_path = 'path/to/experiment/results/visualization';

Visualize the i-th 3D scene interacively by

python utils/visualize.py --result_path out/total3d/[start_time]/visualization --sequence_id [i]

or save the 3D detection result and rendered scene mesh by

python utils/visualize.py --result_path out/total3d/[start_time]/visualization --sequence_id [i] --save_path []

In case you do not have a screen:

python utils/visualize.py --result_path out/total3d/[start_time]/visualization --sequence_id [i] --save_path [] --offscreen

If nothing goes wrong, you should get results like:

724_bbox

724_recon

Visualize the detection results from a third person view with our modified matlab script from Coop:

external/cooperative_scene_parsing/evaluation/vis/show_result.m

Before running the script, please specify the following parameters:

SUNRGBD_path = 'path/to/SUNRGBD';
save_root = 'path/to/save/the/detection/results';
paths = {
    {'path/to/save/detection/results', 'path/to/experiment/results/visualization'}, ...
    {'path/to/save/gt/boundingbox/results'}
};
vis_pc = false; % or true, if you want to show cloud point ground truth
views3d = {'oblique', 'top'}; % choose prefered view
dosave = true; % or false, please place breakpoints to interactively view the results.

If nothing goes wrong, you should get results like:

724_oblique_3d

About the testing speed

Thanks to the simplicity of LIEN+LDIF, the pretrain takes only about 8 hours on a 1080Ti.
However, although we used the CUDA kernel of LDIF to optimize the speed,
the file-based interface of the kernel still bottlenecked the mesh reconstruction.
This is the main reason why our method takes much more time in object and scene mesh reconstruction.
If you want speed over mesh quality, please lower the parameter data.marching_cube_resolution in the configuration file.

Citation

If you find our work and code helpful, please consider cite:

@InProceedings{Zhang_2021_CVPR,
    author    = {Zhang, Cheng and Cui, Zhaopeng and Zhang, Yinda and Zeng, Bing and Pollefeys, Marc and Liu, Shuaicheng},
    title     = {Holistic 3D Scene Understanding From a Single Image With Implicit Representation},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2021},
    pages     = {8833-8842}
}

We thank the following great works:

Total3DUnderstanding for their well-structured code. We construct our network based on their well-structured code.
Coop for their dataset. We used their processed dataset with 2D detector prediction.
LDIF for their novel representation method. We ported their LDIF decoder from Tensorflow to PyTorch.
Graph R-CNN for their scene graph design. We adopted their GCN implemention to construct our SGCN.
Occupancy Networks for their modified version of mesh-fusion pipeline.

If you find them helpful, please cite:

@InProceedings{Nie_2020_CVPR,
author = {Nie, Yinyu and Han, Xiaoguang and Guo, Shihui and Zheng, Yujian and Chang, Jian and Zhang, Jian Jun},
title = {Total3DUnderstanding: Joint Layout, Object Pose and Mesh Reconstruction for Indoor Scenes From a Single Image},
booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2020}
}
@inproceedings{huang2018cooperative,
  title={Cooperative Holistic Scene Understanding: Unifying 3D Object, Layout, and Camera Pose Estimation},
  author={Huang, Siyuan and Qi, Siyuan and Xiao, Yinxue and Zhu, Yixin and Wu, Ying Nian and Zhu, Song-Chun},
  booktitle={Advances in Neural Information Processing Systems},
  pages={206--217},
  year={2018}
}	
@inproceedings{genova2020local,
    title={Local Deep Implicit Functions for 3D Shape},
    author={Genova, Kyle and Cole, Forrester and Sud, Avneesh and Sarna, Aaron and Funkhouser, Thomas},
    booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
    pages={4857--4866},
    year={2020}
}
@inproceedings{yang2018graph,
    title={Graph r-cnn for scene graph generation},
    author={Yang, Jianwei and Lu, Jiasen and Lee, Stefan and Batra, Dhruv and Parikh, Devi},
    booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
    pages={670--685},
    year={2018}
}
@inproceedings{mescheder2019occupancy,
  title={Occupancy networks: Learning 3d reconstruction in function space},
  author={Mescheder, Lars and Oechsle, Michael and Niemeyer, Michael and Nowozin, Sebastian and Geiger, Andreas},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={4460--4470},
  year={2019}
}

GitHub

https://github.com/chengzhag/Implicit3DUnderstanding

Holistic 3D Scene Understanding from a Single Image with Implicit Representation

Implicit3DUnderstanding

Introduction

Install

Demo

Data preparation

Preprocess SUN-RGBD data

Preprocess Pix3D data

Training and Testing

Pretraining

Joint finetune

Testing

About the testing speed

Citation

GitHub

John

Jupyter Notebooks in the terminal

A monocular-based 3D object detection method

Implicit3DUnderstanding

Introduction

Install

Demo

Data preparation

Preprocess SUN-RGBD data

Preprocess Pix3D data

Training and Testing

Pretraining

Joint finetune

Testing

About the testing speed

Citation

GitHub

Jupyter Notebooks in the terminal

A monocular-based 3D object detection method

You might also like...