[Paper Review] Flickr1024: A Large-Scale Dataset for Stereo Image Super-Resolution

Yingqian Wang, Longguang Wang+3 more|arXiv (Cornell University)|Mar 15, 2019

Advanced Image Processing Techniques17 references4 citations

TL;DR

This paper introduces Flickr1024, a large-scale stereo super-resolution dataset comprising 1,024 high-quality, diverse real-world image pairs collected from Flickr with user permissions. The dataset significantly improves stereo SR performance by reducing overfitting and boosting PSNR/SSIM across multiple benchmarks, outperforming KITTI and Middlebury datasets in cross-dataset evaluations using state-of-the-art models like PASSRnet and StereoSR.

ABSTRACT

With the popularity of dual cameras in recently released smart phones, a growing number of super-resolution (SR) methods have been proposed to enhance the resolution of stereo image pairs. However, the lack of high-quality stereo datasets has limited the research in this area. To facilitate the training and evaluation of novel stereo SR algorithms, in this paper, we present a large-scale stereo dataset named Flickr1024, which contains 1024 pairs of high-quality images and covers diverse scenarios. We first introduce the data acquisition and processing pipeline, and then compare several popular stereo datasets. Finally, we conduct crossdataset experiments to investigate the potential benefits introduced by our dataset. Experimental results show that, as compared to the KITTI and Middlebury datasets, our Flickr1024 dataset can help to handle the over-fitting problem and significantly improves the performance of stereo SR methods. The Flickr1024 dataset is available online at: https://yingqianwang.github.io/Flickr1024.

Motivation & Objective

• To address the lack of large-scale, high-quality stereo super-resolution datasets with diverse real-world scenarios.
• To facilitate the training and evaluation of deep learning-based stereo SR methods beyond existing datasets limited to driving or lab environments.
• To investigate whether a larger, more diverse dataset can improve generalization and reduce overfitting in stereo SR models.
• To provide a comprehensive, publicly available dataset that supports both industrial and academic research in stereo image super-resolution.

Proposed method

• The dataset was constructed by manually collecting 1,024 stereo image pairs from Flickr, ensuring permissions were obtained from original photographers.
• Images were processed through a pipeline that corrected cross-eye alignment to parallel optical axes, ensuring proper stereo geometry for super-resolution tasks.
• A multi-step data processing pipeline included image alignment, resolution standardization, and perceptual quality assessment using metrics like BRISQE, ENIQA, and SR-metric.
• The dataset was partitioned into training, validation, and test sets with balanced distribution across perceptual quality metrics to minimize bias.
• Cross-dataset evaluation was performed using state-of-the-art stereo SR models (StereoSR and PASSRnet) trained on different datasets and tested on all benchmarks.
• Performance was quantitatively evaluated using PSNR and SSIM across multiple datasets, including KITTI, Middlebury, ETH3D, and Flickr1024 itself.

Experimental results

Research questions

• RQ1 Can a large-scale, diverse stereo image dataset improve the generalization performance of deep learning-based stereo super-resolution models?
• RQ2 How does the Flickr1024 dataset compare to existing stereo datasets (KITTI, Middlebury, ETH3D) in terms of image quality, diversity, and resolution?
• RQ3 Does training on Flickr1024 reduce overfitting in stereo SR models compared to training on smaller, less diverse datasets?
• RQ4 To what extent does the Flickr1024 dataset enhance PSNR and SSIM performance across different test sets in cross-dataset evaluations?

Key findings

• The Flickr1024 dataset achieved the highest average PSNR and SSIM scores across all test sets in cross-dataset evaluations, outperforming models trained on KITTI2015 and Middlebury datasets.
• Models trained on Flickr1024 showed consistently improving performance with increasing training epochs, indicating reduced overfitting, while models trained on KITTI2015 exhibited performance degradation after a certain point.
• The Flickr1024 dataset improved average PSNR by 1.04 dB over KITTI2015 and 0.58 dB over Middlebury when evaluated on the same test sets.
• The dataset demonstrated superior perceptual image quality, with the highest ENIQA score (0.065) and strong BRISQE (19.40) and SR-metric (7.12) values, indicating high visual quality and texture richness.
• Statistical comparisons showed that Flickr1024 has a balanced distribution of image quality metrics across training, validation, and test sets, minimizing data leakage and bias.
• The dataset's diverse real-world scenarios, including natural scenes and daily photography, make it highly suitable for real-world deployment of stereo SR models in mobile devices.