360 VISTA-SR Dataset

Overview

We provide a dataset containing 200 360-degree videos, predominantly sourced from YouTube and ODV360 (Link) characterized by high quality and resolution (4K and 2K) in ERP format. All videos are licensed under Creative Commons Attribution (reuse allowed), and our dataset is exclusively designed for academic and research purposes. The video dataset encompasses various content characteristics, including outdoor and indoor scenes, as well as high motion sport contents. Each video consists of 100 frames. The dataset is partitioned into 160 videos for training, 20 for validation, and 20 for testing. Note that additional external content can be incorporated for training.

Get dataset from here!

Dataset information

Source content distribution in paired feature space with corresponding convex hulls: Temporal Information(TI) against Spatial Information(SI). VQEG implementation

Source content distribution in paired feature space with corresponding convex hulls: Spatial complexity(E) against Temporal complexity(H). VCA implementation

Results of example baseline models: SWIN and FSRCNN

The table illustrates the WS-PSNR performance and run time of three baseline models on the 360 VISTA-SR validation set. Across both x2 and x4 tracks, FSRCNN emerges as the top-performing model based on our scoring criteria. Despite SwinIR exhibiting superior quality, FSRCNN's faster run time provides it with a competitive advantage. Therefore, the optimal model is one that effectively balances quality and complexity.

Note on computational specifications: The results presented herein were obtained using a desktop computer equipped with an Intel® Xeon 8280 CPU @ 2.70GHz × 56, 128GB RAM, and a 48GB VRAM NVIDIA RTX 6000 Ada graphics card.