State-Space Models for No-Reference Super-Resolution Video Quality Assessment
researchvideo-qualitydeep-learning
Used an array of super-resolution algorithms (image and video) to generate 13,000 up-sampled video sequences with varying quality levels (controlled by the up-sampling scale, and quality of algorithm used). Randomly sampled spatio-temporal patches from these sequences and generated pseudo-labels by computing VMAF scores.
These patches were used in a contrastive pre-training scheme for a no-reference quality assessment metric.
Inference on BVI-SR
| Model | SROCC↑ | PLCC↑ | KRCC↑ |
|---|---|---|---|
| LPIPS | 0.757 | 0.752 | 0.553 |
| FAST-VQA | 0.738 | 0.782 | 0.541 |
| CONVIQT | 0.779 | 0.824 | 0.563 |
| COVER | 0.816 | 0.860 | 0.630 |
| MambaVQA (Ours) | 0.820 | 0.836 | 0.612 |
Performance of no-reference VQA methods on BVI-SR (Mackin et al., 2018). Bold: best, underlined: second best.