IISA (ICCV 2025)

Image Intrinsic Scale Assessment: Bridging the Gap Between Quality and Resolution

Vlad Hosu*, Lorenzo Agnolucci*, Daisuke Iso, Dietmar Saupe,

* Equal contribution.

This is the official repository of the ICCV 2025 paper "Image Intrinsic Scale Assessment: Bridging the Gap Between Quality and Resolution".

Overview

Abstract:
Image Quality Assessment (IQA) measures and predicts perceived image quality by human observers. Although recent studies have highlighted the critical influence that variations in the scale of an image have on its perceived quality, this relationship has not been systematically quantified. To bridge this gap, we introduce the Image Intrinsic Scale (IIS), defined as the largest scale where an image exhibits its highest perceived quality. We also present the Image Intrinsic Scale Assessment (IISA) task, which involves subjectively measuring and predicting the IIS based on human judgments. We develop a subjective annotation methodology and create the IISA-DB dataset, comprising 785 image-IIS pairs annotated by experts in a rigorously controlled crowdsourcing study with verified reliability. Furthermore, we propose WIISA (Weak-labeling for Image Intrinsic Scale Assessment), a strategy that leverages how the IIS of an image varies with downscaling to generate weak labels. Experiments show that applying WIISA during the training of several IQA methods adapted for IISA consistently improves the performance compared to using only ground-truth labels.

Citation

If you find our work useful, please cite our paper:

@inproceedings{hosu2025image,
  title={Image Intrinsic Scale Assessment: Bridging the Gap Between Quality and Resolution},
  author={Hosu, Vlad and Agnolucci, Lorenzo and Iso, Daisuke and Saupe, Dietmar},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year={2025}
}

IISA-DB Dataset

We introduce the IISA-DB dataset, which contains 785 image-IIS pairs annotated by experts in a rigorously controlled crowdsourcing study with verified reliability. The dataset is available for download from here. In our experiments, we randomly partition the dataset into 70% training, 10% validation, and 20% testing sets over 10 different splits. The splits are available in datasets/splits.

The dataset class is implemented in pyiqa/data/iisadb_dataset.py. The code supports loading the dataset and generating weak labels for training using the WIISA strategy proposed in our paper.

Getting Started

Installation

git clone https://github.com/SonyResearch/IISA.git
cd IISA
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
python setup.py develop

Data Preparation

After downloading the IISA-DB dataset, ensure that the folder structure matches the following:

├── IISA-DB
│   annotations.csv

│   ├── images
|   |   ├── [374612387.jpg | 6246977871.jpg | 9470847928.jpg | ...]

Inference with Pre-trained Models

Thanks to torch.hub, you can leverage our pre-trained checkpoints for inference without the need to clone our repo. The code snippet below shows how to run inference with models pre-trained on the IISA-DB dataset. You have to change <METRIC_NAME> with one of the following:

• arniqa: ARNIQA
• clipiqa: CLIPIQA
• contrique: CONTRIQUE
• dbcnn: DBCNN
• qualiclip: QualiCLIP
• topiq: TOPIQ

import pyiqa
from PIL import Image
from torchvision.transforms import ToTensor
import torch

device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")

# Load the pre-trained model
model = torch.hub.load(repo_or_dir="SonyResearch/IISA", model="<METRIC_NAME>", device=device)    # The model is already in evaluation mode

# Option 1: Inference using image tensor
img = Image.open('assets/example_image.jpg').convert('RGB')
img_tensor = ToTensor()(img).unsqueeze(0).to(device)    # The models expect non-normalized batched image tensors

with torch.no_grad():
    pred = model(img_tensor)

print('Predicted IIS from tensor: ', pred.item())

# Option 2: Inference using image path
pred = model('assets/example_image.jpg')
print('Predicted IIS from path: ', pred.item())

Training

To train a model on the IISA-DB dataset, you have to first modify the dataset_path variable in the corresponding config file under the configs/ directory to point to the location of the dataset on your machine, such as /home/IISA-DB. Optionally, you can also modify the logger.wandb variable to enable logging with Weights & Biases.

Then, run the following command to cross-validate the model on 10 splits, using the checkpoint with the best validation performance to evaluate the model on the test set:

python pyiqa/train_nsplits.py -opt configs/train_<MODEL_NAME>_iisadb.yml

where <MODEL_NAME> can be one of the following: ARNIQA, CLIPIQA, CONTRIQUE, DBCNN, QualiCLIP and TOPIQ.

Alternatively, you can train the model on a single split by running:

python train.py -opt configs/train_<MODEL_NAME>_iisadb.yml

Acknowledgements

Our code is based on the IQA-PyTorch library. We thank the authors for their contribution.