Flickr1024: A Large-Scale Dataset for Stereo Image Super-Resolution

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.

Learning Parallax Attention for Stereo Image Super-Resolution

“…1: Visual results achieved by PASSRnet [2], [3], SRRes-Net+SAM [4], and our iPASSR network for 4×SR. These results are achieved on scene Test 0002 of the Flickr1024 dataset [7]. Our method can recover more faithful and stereoconsistent details than other methods.…”

“…During the test phase, we followed [1]- [4] to use 5 images from the Middlebury dataset [10], 20 images from the KITTI 2012 dataset [8] and 20 images from the KITTI 2015 dataset [9]. Furthermore, we used the test set provided by the Flickr1024 dataset [7] for additional evaluation.…”

“…That is, the converted features in biPAM are directly concatenated with the original features on the target side without using Eqs. (6) and (7). Meanwhile, the photometric loss, cycle loss, and stereo consistency loss are applied equally to all spatial locations without considering occlusions.…”

“…As demonstrated in [12]- [14], a deep and large SR model enables rich and hierarchical feature representation (i.e., enhance intra-view information exploitation) and thus can significantly boost the SR performance. Differently, our [9] and Flickr1024 [7] datasets for 4×SR.…”

Light Field Image Super-Resolution Using Deformable Convolution

“…In this section, we compare our NUDT dataset to several popular LF datasets [16], [69]- [72]. Following [92], we use four no-reference image quality assessment (NRIQA) metrics (i.e, BRISQUE [64], NIQE [65], CEIQ [66], ENIQA [67]) to evaluate the perceptual quality of the center-view images of these datasets. Besides, we also use a no-reference LF quality assessment metric (i.e., NRLFQA [68]) to evaluate TABLE I: Main characteristics of several popular LF datasets.…”