The Graphics-LPIPS metric is an extension of the LPIPS metric originally designed for images and perceptual similarity tasks, which we adapted for 3D graphics and quality assessment tasks. Graphics-LPIPS employs CNNs with learning linear weights on top, fed with reference and distorted patches of rendered images of 3D models. The overall quality of the 3D model is derived by averaging local patch qualities
This project is the implementation of our paper: Textured Mesh Quality Assessment: Large-Scale Dataset and Deep Learning-based Quality Metric. Yana Nehmé, Johanna Delanoy, Florent Dupont, Jean-Philippe Farrugia, Patrick Le Callet, Guillaume Lavoué
- Install PyTorch 1.0+ and torchvision fom http://pytorch.org
pip install -r requirements.txt- Clone this repo:
git clone https://github.com/YanaNEHME/Graphics-LPIPS
cd Graphics-LPIPSGraphics-LPIPS predicts a quality score ranging between [0,1]. The higher the Graphics-LPIPS values, the more different the patches are.
Example of script to compute the distance between a reference patch (p0) and a distorted patch (p1).
You can pass in parameter (-m or --modelpath) the path of the network to use, usually located in the ./checkpoints directory.
To use the GPU, turn on the `--use_gpu`` parameter.
The output is the predicted quality score of the distorted patch.
python GraphicsLpips_2imgs.py -p0 imgs/ex_ref.png -p1 imgs/ex_p0.png --use_gpu
Example of script to compute the quality of a set of distorted 3D graphics. The snapshots of the 3D models are patchified (divided into small patches). The number of patches obtained for each model is stored in a csv file. Graphics-LPIPS estimates the quality locally (i.e. per patch), then the global quality score of the model is computed as the average of the local patch qualities.
python GraphicsLpips_csvFile.py -m './checkpoints/GraphicsLPIPS_FinalNetwork/latest_net_.pth' --use_gpu
Graphics-LPIPS was trained and tested on a challenging dataset of 3000 textured meshes. The dataset was generated from 55 source models corrupted by combinations of 5 types of compression-based distortions applied on the geometry, texture mapping and texture image of the meshes. The stimuli were annotated in a large-scale subjective experiment in crowdsourcing, based on the DSIS method. Thus, each stimulus is associated with a subjective quality score a.k.a Mean Opinion Score (MOS).
Download the dataset (3D models and their distorted versions, snapshots of the stimuli taken from their main viewpoint, subjective scores)
Graphics-LPIPS is designed for 3D graphics and quality assessment tasks. To predict the overall quality of a stimulus, we modified the original LPIPS metric so that: (1) the small network (G), trained on top, suits MOS scores instead of preference scores, and (2) the optimization (the loss computation) is done per image (instead of patch-wise).
We employed the pre-trained AlexNet network with its fixed weights and learnt the weights of a linear layer on top.
To train our model, we considered images of the 3D models taken from their main viewpoint to which we associated the MOS scores.
The images were divided into patches of size 64x64.
See scripts train.py and ./scripts/train_metric.txt for an example of training and testing the metric.
The scripts will train a model on randomly sampled patches of stimulus images from the training set, for --nepoch + --nepoch_decay epochs.
As the distances computed for patches of the same image are combined for the loss calculation, the patches of the same image can not be distributed over different batches.
Thus, each batch was made to contain --nInputImg images, each represented by --npatches randomly sampled patches. This
results in a batch size of --nInputImg x --npatches patches.
The backpropagated error is the average loss over the images in a batch.
During training, patches are randomly sampled every epoch to ensure that as many different image patches as possible are used in training. 80% of the stimuli in the dataset are used for training and 20% for testing. Subjective scores were scaled between [0,1] - 0: imperceptible distortion (highest quality), 1: very annoying distortion (lowest quality).
Training will add a subdirectory in the checkpoints directory.
Download the dataset patchified (used to train and test the metric), and unzip it into directory ./dataset
This work was supported by French National Research Agency as part of ANR-PISCo project (ANR-17-CE33-0005).
Yana Nehmé, Johanna Delanoy, Florent Dupont, Jean-Philippe Farrugia, Patrick Le Callet, Guillaume Lavoué, Textured mesh quality assessment: Large-scale dataset and deep learning-based quality metric, ACM Transactions on Graphics, to be presented at SIGGRAPH 2023.
The Graphics-LPIPS metric is Copyright of University of Lyon, 2022.
It is distributed under the Mozilla Public License v. 2.0. (refer to the accompanying file LICENSE-MPL2.txt or a copy at http://mozilla.org/MPL/2.0/)