Motivation

Figure 1 Video contents possess diverse motion cues and various restoration difficulties.

Ø We pioneer a novel representation for video modeled as instances,

events, and scenes, providing both global semantics and instance-

specific semantics to boost the performance of video super-resolution.

Ø We propose a Semantics-Powered Attention Cross-Embedding

block to bridge semantic priors and pixel-level features, being aware

of the restored contents.

Ø We further design Instance-Specific Semantic Embedding Encoder

1. Features of supporting frames are first coarsely

Semantic Extractor decouples pixel-based video into instances, events,

and scenes, and Pixel Enhancer refines the original pixels and

generates absent information with the guidance of the semantic priors.

Methodology

Ø With the aim of improving the inter-frame alignment, we formulate a

semantic-guided instance-centric alignment schema.