Home

Awesome

Shape and Material Capture at Home, CVPR 2021.

Daniel Lichy, Jiaye Wu, Soumyadip Sengupta, David Jacobs

<!--<figure class="image"> <img src="./media/teaser2.png"> <figcaption>{bare-bones capture setup}</figcaption> </figure>--> <p> <img src="./media/teaser2.png" alt> <em>A bare-bones capture setup</em> </p>

Overview

This is the official code release for the paper Shape and Material Capture at Home. The code enables you to reconstruct a 3D mesh and Cook-Torrance BRDF from one or more images captured with a flashlight or camera with flash.

We provide:

Dependencies

This project uses the following dependencies:

The easiest way to run the code is by creating a virtual environment and installing the dependences with pip e.g.

# Create a new python3.8 environment named py3.8
virtualenv py3.8 -p python3.8

# Activate the created environment
source py3.8/bin/activate

#upgrade pip
pip install --upgrade pip

# To install dependencies 
python -m pip install -r requirements.txt
#or
python -m pip install -r requirements_no_exr.txt


Capturing you own dataset

Multi-image captures

The video below shows how to capture the (up to) six images for you own dataset. Angles are approximate and can be estimated by eye. The camera should be approximately 1 to 4 feet from the object. The flashlight should be far enough from the object such that the entire object is in the illumination cone of the flashlight.

We used this flashlight, but any bright flashlight should work. We used this tripod which comes with a handy remote for iPhone and Android.

<img src="./media/capture_process_gif.gif" width="960" height="540">

Please see the Project Page for a higher resolution version of this video.

Example reconstructions:

<p float="left"> <img src="./media/bear_cup/images_s5_b0_ch6.png" width="180" height="180"> <img src="./media/black_skull/images_s5_b0_ch6.png" width="180" height="180"> <img src="./media/white_skull/images_s5_b0_ch6.png" width="180" height="180"> <img src="./media/hat/images_s5_b0_ch6.png" width="180" height="180"> <br> <img src="./media/bear_cup/object_motion_video_gif.gif" width="180" height="180"> <img src="./media/black_skull/object_motion_video_gif.gif" width="180" height="180"> <img src="./media/white_skull/object_motion_video_gif.gif" width="180" height="180"> <img src="./media/hat/object_motion_video_gif.gif" width="180" height="180"> </p>

Single image captures

Our network also provides state-of-the-art results for reconstructing shape and material from a single flash image.

Examples captured with just an iPhone with flash enabled in a dim room (complete darkness is not needed):

<p float="left"> <img src="./media/single_skull/images_s5_b0_ch6.png" width="200" height="200"> <img src="./media/single_starwars_cup/images_s5_b0_ch6.png" width="200" height="200"> <br> <img src="./media/single_skull/object_motion_video_gif.gif" width="200" height="200"> <img src="./media/single_starwars_cup/object_motion_video_gif.gif" width="200" height="200"> </p>

Mask Making

For best performance you should supply a segmentation mask with your image. For our paper we used https://github.com/saic-vul/fbrs_interactive_segmentation which enables mask making with just a few clicks.

Normal prediction results are reasonable without the mask, but integrating normals to a mesh without the mask can be challenging.

Test RecNet on the DiLiGenT dataset

# Download and prepare the DiLiGenT dataset
sh scripts/prepare_diligent_dataset.sh

# Test on 3 DiLiGenT images from the front, front-right, and front-left
# if you only have CPUs remove the --gpu argument
python eval_diligent.py results_path --gpu

# To test on a different subset of DiLiGenT images use the argument --image_nums n1 n2 n3 n4 n5 n6
# where n1 to n6 are the image indices of the right, front-right, front, front-left, left, and above
# images, respectively. For images that are no present set the image number to -1
# e.g to test on only the front image (image number 51) run
python eval_diligent.py results_path --gpu --image_nums -1 -1 51 -1 -1 -1 

Test on our dataset/your own dataset

The easiest way to test on you own dataset and our dataset is to format it as follows:

dataset_dir:

Where .ext is the image extention e.g. .png, .jpg, .exr

For an example of formating your own dataset please look in data/sample_dataset

Then run:

python eval_standard.py results_path --dataset_root path_to_dataset_dir --gpu

# To test on a sample of our dataset run
python eval_standard.py results_path --dataset_root data/sample_dataset --gpu

Download our real dataset

Our real dataset can be downloaded here (56MB). To test our network on it, first unzip it, then run with the command above, setting "dataset_root" to the path of the unzipped data.

Integrating Normal Maps and Producing a Mesh

We include a script to integrate normals and produce a ply mesh with per vertex albedo and roughness.

After running eval_standard.py or eval_diligent.py there with be a file results_path/images/integration_data.csv Running the following command with produce a ply mesh in results_path/images/sample_name/mesh.ply

python integrate_normals.py results_path/images/integration_data.csv --gpu

This is the most time intensive part of the reconstruction and takes about 3 minutes to run on GPU and 5 minutes on CPU.

Training

To train RecNet from scratch:

python train.py log_dir --dr_dataset_root path_to_dr_dataset --sculpt_dataset_root path_to_sculpture_dataset --gpu

Download the training data

Our network is trained on two synthetic datasets available here: synthetic_sculpture_dataset (23GB) and synthetic_DR_dataset (45GB)

synthetic_sculpture_dataset uses meshes from the sculpture dataset available at https://www.robots.ox.ac.uk/~vgg/data/SilNet/ synthetic_DR_dataset uses random meshes available at https://github.com/zexiangxu/Deep-Relighting

FAQ

Q1: What should I do if I have problem running your code?

Citation

If you find this code or the provided models useful in your research, please cite it as:

@inproceedings{lichy_2021,
  title={Shape and Material Capture at Home},
  author={Lichy, Daniel and Wu, Jiaye and Sengupta, Soumyadip and Jacobs, David W.},
  booktitle={CVPR},
  year={2021}
}

Acknowledgement

Code used for downloading and loading the DiLiGenT dataset is adapted from https://github.com/guanyingc/SDPS-Net