NVS-SQA

Official implementation of "NVS-SQA: Exploring Self-Supervised Quality Representation Learning for Neurally Synthesized Scenes without References" arxiv) (accepted by TPAMI).

New: you can now install it via pip!

Installation

You can install the package directly from PyPI:

pip install nvs_sqa

Usage

The package can be used programmatically: The program allows you to utilize quality representations for downstream tasks and obtain quality scores from the extracted features. You can download the examples folder here. The quality scores are in format of JOD score, mostly negative values (offset by reference quality), with higher scores indicating better quality. The basic logic and output format can be found in Section Generating No-Reference Quality Representations with the Pretrained Model. Note: you can normalize the prediction on your dataset for more relevant scores. Here's an example of how to generate quality features and scores:

import nvs_sqa

# Initialize the QualityAssessor
qa = nvs_sqa.QualityAssessor()

# Define the evaluation folder path containing NSS directories
eval_folder = "./examples"

# Generate quality features
all_feats, save_path = qa.generate_quality_features(eval_folder, verbose=True)

# Compute quality scores
quality_scores = qa.generate_quality_scores(all_feats, verbose=True)

# The quality scores are in format of JOD score, mostly negative values (offset by reference quality), with higher scores indicating better quality.

print(f"Features saved to: {save_path}")

Visualization of Some Examples:

Requirements

numpy==1.24.1 opencv_contrib_python==4.8.0.76 opencv_python==4.8.0.76 Pillow==10.2.0 pyfvvdp==1.2.0 scikit_learn==1.3.0 scikit_video==1.1.11 scipy==1.12.0 tabulate==0.9.0 torch==2.0.1+cu118 torchsummary==1.5.1 torchvision==0.15.2+cu118

Evaluate the Pretrained Model

Execute python3 evaluate.py to obtain datasetwise and scenewise experiment results.

One Linear Model for Three Datasets

Performance on Each Dataset

Metric	Fieldwork	LLFF	Lab
SRCC	0.9111	0.7000	0.7000
PLCC	0.8828	0.6441	0.6783

Performance on Each Scene

Scene	Animals	Bears	CD-Occ.	Cd_occlusion_intr	Dinosaur	Elephant	Fern	Flower	Fortress	Giraffe	Glass	Glossy_animals_extr	Horns	Leaves	Leopards	Metal	Naiad-Sta.	Orchids	Puccini_statue	Room	Toys	Trex	Vespa	Whale
SRCC	0.5000	0.9000	0.8000	0.1000	0.8000	1.0000	0.0000	0.9000	0.7000	0.9000	0.9000	0.9000	0.9000	0.7000	0.9000	0.8000	0.9000	0.8000	0.8000	0.7000	0.9000	0.9000	1.0000	1.0000
PLCC	0.2545	0.7628	0.8159	0.0470	0.8035	0.9944	-0.3679	0.7713	0.6554	0.9361	0.9735	0.9432	0.7288	0.8445	0.9395	0.7438	0.8370	0.9498	0.7087	0.7369	0.9704	0.8341	0.9901	0.9734

Generating No-Reference Quality Representations with the Pretrained Model

To generate quality features for example Neurally Synthesized Scenes (NSS), follow these steps:

1. Run the Application: Execute python3 app.py. This script processes each NSS folder within the ./examples directory.
2. Feature Generation and Saving: The script iterates through all NSS folders in ./examples, generates quality features for each, and saves these features in ./examples/output_features.npz. output_features.npz is in format of {<nss_name>: 384-dim rep., ...}. (You can use np.load() to load it.)
3. Example NSS Provided: Within the examples folder, we include two NSS examples: one generated by GNT-Cross-scene and the other by Plenoxel.
4. Quality Score Output: app.py also calculates and outputs quality scores based on the generated features applying a ridge linear regression model on training segements of Fieldwork, Lab, and LLFF datasets.

Important Notes:

• JOD Scoring Format: The JOD score, which is the adopted scoring format, features primarily negative values (offset by reference quality), with higher scores indicating better quality.
• Relevance of JOD Scores: According to the dataset authors for Fieldwork and Lab, JOD scores are more meaningful within the same scene, suggesting that cross-scene comparisons of JOD scores may not provide meaningful insights.

Training the Model via Self-Supervised Learning

Our model was developed using a dataset sourced from Nerfstudio. For detailed instructions on obtaining this dataset, visit Nerfstudio's Dataset Quickstart.

Installation of Nerfstudio

Follow the installation guide provided on the Nerfstudio Installation Page to set up Nerfstudio on your system.

Dataset Generation for Self-Supervised Learning

After installing Nerfstudio, execute the following steps to automatically prepare the dataset for self-supervised learning:

1. Download the source data: ns-download-data nerfstudio --capture-name all
2. Unzip each scene into the datasets/nerfstudio directory.
3. Run the preprocessing script: python3 preprocess.py

Note: Generating the unlabeled Neurally Synthesized Scenes (NSS) may take several days. Adjust the nerfstudio_scenes and nerfstudio_methods variables in the consts.py file to select specific scenes and NeRF methods for generation. The approximate processing times per method per scene are as follows:

• splatfacto: ~5 mins
• instant-ngp: ~20 mins
• nerfacto: ~15 mins
• tensorf: ~20 mins
• kplanes: ~30 mins
• nerfacto-huge: 3 hrs

Training Process

Modify parameters such as version, batch_size, and seq_model_config in train.py. Execute the training script with python3 train.py.

Model Evaluation

After training, update the version and epo_offset in evaluate.py to match your trained model's version and epoch number. Evaluate the pretrained model by running python3 evaluate.py.

To Cite

@article{qu2025nvs,
  title={NVS-SQA: Exploring Self-Supervised Quality Representation Learning for Neurally Synthesized Scenes without References},
  author={Qu, Qiang and Shen, Yiran and Chung, Yuk Ying and Cai, Weidong and Chen, Xiaoming and Liu, Tongliang},
  journal={arXiv preprint arXiv:2501.06488},
  year={2025}
}

@article{qu2024nerf,
  title={NeRF-NQA: No-Reference Quality Assessment for Scenes Generated by NeRF and Neural View Synthesis Methods},
  author={Qu, Qiang and Liang, Hanxue and Chen, Xiaoming and Chung, Yuk Ying and Shen, Yiran},
  journal={IEEE Transactions on Visualization and Computer Graphics},
  year={2024},
  publisher={IEEE}
}

@inproceedings{liang2024perceptual,
  title={Perceptual Quality Assessment of NeRF and Neural View Synthesis Methods for Front-Facing Views},
  author={Liang, Hanxue and Wu, Tianhao and Hanji, Param and Banterle, Francesco and Gao, Hongyun and Mantiuk, Rafal and {\"O}ztireli, Cengiz},
  booktitle={Computer Graphics Forum},
  volume={43},
  number={2},
  pages={e15036},
  year={2024},
  organization={Wiley Online Library}
}