1 Introduction

https://github.com/tsingqguo/misf

https://arxiv.org/pdf/2203.06304.pdf

MISF: Multi-level Interactive Siamese Filtering for High-Fidelity Image Inpainting, cvpr22

Untitled

a) predictive image-level filtering: HxWx3k^2

b) semantic feature-level filtering; c) proposed multi-level filtering.

native generative network-based inpainting or native encoder-decoder (En-Decoder) ⇒ results like b). This paper also contains a L_gan loss.

Untitled

2 Why do we still discuss MISF in cv course?

2.1 A little shocking news for me several years before.

2.1.1 Dynamic filters & deformable convolutions

1.3.2.2.2 Decoupled Dynamic Filter Networks, cvpr21

以前的这些工作，基本都只用在high level semantic app，只在特征层用，只为了更好的特征提取工具。

2.2 Predictive filtering based image restoration

  1. denoising [1, 21], deraining [13], shadow removing [9], and blur synthesis [4] =》只做了image filtering，没做feature filtering。

  2. 3篇系列工作，从deraining [13]，做到简单的和generation method的结合 [12]，到在特征上学、用kernel [MISF, cvpr22]。

  3. [13] EfficientDeRain: Learning Pixel-wise Dilation Filtering for High-Efficiency Single-Image Deraining, AAI 2021

    image.png

    Figure 2: Pipeline of our EfficientDeRain. (a) 标准的pixel-wise image filtering where each pixel is processed by a kernel predicted by a kernel prediction network. (b) 拓展为 pixel-wise dilation filtering 空洞卷积 to handle multi-scale rain streaks.