IC-767BSRNet outperforms RealESRNet on the majority of degradation cases, reversing the ranking obtained from a single random test set

Wenlong Zhang, Xiaohui Li, Xiangyu Chen, Xiaoyun Zhang, Yu Qiao, Xiao-Ming Wu, Chao Dong

SourceSEAL: A Framework for Systematic Evaluation of Real-World Super-Resolution

On a single randomly generated test set (real-DIV2K val), RealESRNet achieves a higher average PSNR than BSRNet (24.93 dB vs 24.77 dB), suggesting RealESRNet is superior. However, under SEAL's systematic evaluation on 100 representative degradation cases, BSRNet achieves AR 0.55 while RealESRNet achieves only AR 0.15, meaning BSRNet outperforms the FSRCNN acceptance line on 55% of cases versus 15% for RealESRNet. Across 20 random test sets, the PSNR difference between the two ranges from -0.22 to 0.18 dB, making the conventional ranking unstable.

Evidence
correlational
Key metric
PSNR on real-DIV2K val: RealESRNet 24.93 dB vs BSRNet 24.77 dB; AR on DIV2K-SE: BSRNet 0.55 vs RealESRNet 0.15; PSNR difference across 20 random sets ranges from -0.22 to 0.18 dB
Caveat
The conventional test set uses a single random sampling of degradations; the SEAL test set uses 100 cluster centers from spectral clustering of 10,000 sampled degradations.
Model
BSRNet, RealESRNet
Concepts
Failure mode
Datasets
DIV2K-SE [eval], DIV2K val [eval]
Related work
BSRGAN [context]
Related findings
IC-766, IC-768
Extraction
automatic-extraction