Author: Satoshi Ikehata (National Institute of Informatics https://satoshi-ikehata.github.io/)
Conference: IEEE/CVF Computer Vision and Pattern Recognition (CVPR), 2023
Arxiv: https://arxiv.org/abs/2303.15724
Abstract: SDM-UniPS is a Scalable, Mask-free, and Universal Photometric Stereo Network. This method can recover remarkably intricate surface normal maps, rivaling the quality of 3D scanners, even when images are captured under unknown, spatially-varying lighting conditions in uncontrolled environments. I provide not only a model for normal map estimation but also one for BRDF parameters (base color, roughness, and metallic), as shown in the appendix. Additionally, I offer the code for creating a video from frames rendered under novel directional light sources using both the normal and BRDF parameters.
What is Universal Photometric Stereo (UniPS)? Unlike existing calibrated and uncalibrated photometric stereo tasks, which assume specific physical lighting models and therefore limit their usability, the solution algorithm of universal photometric stereo is designed to work for objects with diverse shapes and materials under arbitrary lighting variations, without assuming any specific models.
Follow instructions below.
Capture multiple images under various arbitrary lighting conditions with a fixed camera. Unlike existing photometric stereo methods, there is no assumption of directional lighting. However, please ensure that there is sufficient shading variation in all parts of the images. You can find tips for capture in the appendix of the paper.
An additional object mask is useful for cutting out the target object and hiding areas with weak shading variations. It also enhances reconstruction accuracy by providing a powerful cue to indicate boundary constraints to the network. Therefore, I recommend using the mask if your priority is reconstruction accuracy. While this repository does not support mask generation, you can find various public software tools to generate the foreground object mask. For example, Konstantin Sofiiuk's software only requires a few clicks to generate the foreground object mask.
Run main.py to recover the surface normal map and/or BRDF parameter maps (base color, roughness, metallic).
Run relighting.py to render images under novel directional lightings based on recovered normal map and BRDF parameter maps.
To successfully run the universal photometric stereo network, ensure that your system has the following dependencies installed:
- Python 3
- PyTorch
- OpenCV (cv2)
- Einops
- imageio (only for relighting)
The network has been successfully tested on the following configurations:
- Ubuntu 20.04.5 (WSL2) and Windows 11, Python 3.11.3, PyTorch 2.0 with CUDA 11.8
- GPU: Nvidia RTX A6000 (48GB)
To run the universal photometric stereo network, you need shading images and an optional binary object mask. The object should be illuminated under arbitrary lighting sources, but shading variations should be sufficient (weak shading variations may result in poor results).
Organize your test data as follows (prefix "L" and suffix ".data" can be modified in main.py). You can process multiple datasets (A, B, ...) simultaneously, which is convenient if you are evaluating the method on the DiLiGenT benchmark.
YOUR_DATA_PATH
├── A [Suffix (default ".data")]
│ ├── mask.png (optional)
│ ├── [Prefix (default: "L")] imgfile1
│ ├── [Prefix (default: "L")] imgfile2
│ └── ...
├── B [Suffix (default: ".data")]
│ ├── mask.png (optional)
. ├── [Prefix (default: "L")] imgfile1
. ├── [Prefix (default: "L")] imgfile2
└── ...
You can simply try the code using some datasets in my paper from this link. As an alternative option, you can also use the dataset of my past research, UniPS. this link. To generate the object mask, you can simply use public softwares such as this link
Checkpoints of the network parameters for recovering both normal maps and BRDF maps are available here.
To use the pre-trained models, extract them as follows:
YOUR_CHECKPOINT_PATH
├── normal
│ └── nml.pytmodel
└── brdf
└── brdf.pytmodel
To run the test, execute main.py with the following command:
python sdm_unips/main.py --session_name YOUR_SESSION_NAME --test_dir YOUR_DATA_PATH --checkpoint YOUR_CHECKPOINT_PATH
or
python sdm_unips/main.py --session_name YOUR_SESSION_NAME --test_dir YOUR_DATA_PATH --checkpoint YOUR_CHECKPOINT_PATH --scalable
The 'scalable' option is a powerful feature not included in the original paper. At the cost of some accuracy and computational expense, it enables normal and brdf estimation at any resolution with constant GPU memory consumption. For example, in the original implementation, inputting ten 2048x2048 resolution images consumes over 40GB of memory, but using this option reduces consumption to around 10GB. This consumption is proportional to the number of images but does not vary with resolution. The trick is very simple (Basically, it is a similar approach to the Scale-invariant Spatial-light Feature Encoder). If you input images larger than 1024x1024, I highly recommend you to use this option.
If you don't need to recover either of normal map or brdf maps, please use specific options as
python sdm_unips/main.py --session_name YOUR_SESSION_NAME --test_dir YOUR_DATA_PATH --checkpoint YOUR_CHECKPOINT_PATH --target normal
or
python sdm_unips/main.py --session_name YOUR_SESSION_NAME --test_dir YOUR_DATA_PATH --checkpoint YOUR_CHECKPOINT_PATH --target brdf
By default, the maximum number of images to load is set to 10, and the maximum resolution is set to 4096x4096, regardless of the number of images or the resolution in the data directory. If you want to change these numbers, please specify options as follow.
python sdm_unips/main.py --session_name YOUR_SESSION_NAME --test_dir YOUR_DATA_PATH --checkpoint YOUR_CHECKPOINT_PATH --max_image_res 10000 --max_image_num 100
The results will be saved in YOUR_SESSION_NAME/results. You will find a normal map and BRDF maps (base color, roughness, and metallic).
You can also use the provided code (relighting.py) for relighting the object under novel directional lights based on the recovered attributes. Follow the instructions displayed at the end of the prompt to use it. It should look like this.
To output .avi video:
python sdm_unips/relighting.py --datadir ./YOUR_SESSION_NAME/results/OBJECT_NAME.data --format avi
To output .gif video:
python sdm_unips/relighting.py --datadir ./YOUR_SESSION_NAME/results/OBJECT_NAME.data --format gif
If you want to compute the mean angular errors between predicted and ground truth surface normal maps in DiLiGenT benchmark, simply organize the data as in others and put 'Normal_gt.png' in the directory. The code automatically computes the MAE and displays it. It also generates the error map (0 deg. ~ 90 deg.). Please note that there are randomized processes in the estimation framework, and the results change with every prediction.
This repository is for testing purposes only. I do not currently plan to release the training script and training data, unlike my previous projects. Please refrain from sending email requests for these resources.
This project is licensed under the MIT License with a non-commercial clause. This means that you are free to use, copy, modify, and distribute the software, but you are not allowed to use it for commercial purposes.
Copyright (c) [2023] [Satoshi Ikehata]
If you find this repository useful, please consider citing this paper:
@inproceedings{ikehata2023sdmunips,
title={Scalable, Detailed and Mask-free Universal Photometric Stereo},
author={Satoshi Ikehata},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
year={2023}
}