Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

NTIRE 2024 Challenge on Stereo Image Super-Resolution: Methods and Results

Longguang Wang, Yulan Guo, Juncheng Li, Hongda Liu, Yang Zhao, Yingqian Wang, Zhi Jin, Shuhang Gu, Radu Timofte

TL;DR

The NTIRE 2024 Challenge on Stereo Image SR investigates x4 super-resolution for LR stereo pairs under tight computational budgets, emphasizing cross-view information for stereo-consistent reconstruction. It surveys related single-image SR and stereo SR work, details a Flickr1024-based dataset with two degradation tracks, and reports finalist performances that predominantly fuse Transformer backbones with parallax-attention-based cross-view fusion under strict size and MAC constraints. Top solutions leverage efficient cross-view interaction (PAM, SCAM, MCAM) and lightweight convolutional blocks (depthwise/NAF variants, Mamba-inspired ideas) to balance accuracy and efficiency, establishing a new benchmark for resource-constrained stereo SR. The results guide practical design choices for real-world stereo SR systems and highlight the continued importance of cross-view cues and efficient architectures.

Abstract

This paper summarizes the 3rd NTIRE challenge on stereo image super-resolution (SR) with a focus on new solutions and results. The task of this challenge is to super-resolve a low-resolution stereo image pair to a high-resolution one with a magnification factor of x4 under a limited computational budget. Compared with single image SR, the major challenge of this challenge lies in how to exploit additional information in another viewpoint and how to maintain stereo consistency in the results. This challenge has 2 tracks, including one track on bicubic degradation and one track on real degradations. In total, 108 and 70 participants were successfully registered for each track, respectively. In the test phase, 14 and 13 teams successfully submitted valid results with PSNR (RGB) scores better than the baseline. This challenge establishes a new benchmark for stereo image SR.

NTIRE 2024 Challenge on Stereo Image Super-Resolution: Methods and Results

TL;DR

The NTIRE 2024 Challenge on Stereo Image SR investigates x4 super-resolution for LR stereo pairs under tight computational budgets, emphasizing cross-view information for stereo-consistent reconstruction. It surveys related single-image SR and stereo SR work, details a Flickr1024-based dataset with two degradation tracks, and reports finalist performances that predominantly fuse Transformer backbones with parallax-attention-based cross-view fusion under strict size and MAC constraints. Top solutions leverage efficient cross-view interaction (PAM, SCAM, MCAM) and lightweight convolutional blocks (depthwise/NAF variants, Mamba-inspired ideas) to balance accuracy and efficiency, establishing a new benchmark for resource-constrained stereo SR. The results guide practical design choices for real-world stereo SR systems and highlight the continued importance of cross-view cues and efficient architectures.

Abstract

This paper summarizes the 3rd NTIRE challenge on stereo image super-resolution (SR) with a focus on new solutions and results. The task of this challenge is to super-resolve a low-resolution stereo image pair to a high-resolution one with a magnification factor of x4 under a limited computational budget. Compared with single image SR, the major challenge of this challenge lies in how to exploit additional information in another viewpoint and how to maintain stereo consistency in the results. This challenge has 2 tracks, including one track on bicubic degradation and one track on real degradations. In total, 108 and 70 participants were successfully registered for each track, respectively. In the test phase, 14 and 13 teams successfully submitted valid results with PSNR (RGB) scores better than the baseline. This challenge establishes a new benchmark for stereo image SR.
Paper Structure (34 sections, 6 equations, 27 figures, 2 tables)

This paper contains 34 sections, 6 equations, 27 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Single Image SR
  4. Stereo Image SR
  5. NTIRE 2024 Challenge
  6. Dataset
  7. Tracks
  8. Challenge Phases
  9. Challenge Results
  10. Track 1: Fidelity & Bicubic Degradation
  11. Track 2: Fidelity & Realistic Degradation
  12. Summary
  13. Challenge Methods and Teams
  14. Davinci - Track 1$^\bigstar$, Track 2$^\bigstar$
  15. HiSSR - Track 1$^\bigstar$
  16. ...and 19 more sections

Figures (27)

  • Figure 1: Comparison between different ensemble strategies. (a) Multi-model ensemble. (b) Data ensemble. (c) The proposed model ensemble. Rectangles with different colors represent different model parameters.
  • Figure 2: Davinci: The structure of the proposed SwinFIRSSR.
  • Figure 3: HiSSR: The structure of the proposed RISSR.
  • Figure 4: MiVideoSR: The structure of the proposed HCASSR.
  • Figure 5: BUPTMM: The structure of the proposed Efficient CVHSSR.
  • ...and 22 more figures