【WACV 2021 Oral】Attentional Feature Fusion

Abstract

We present a multi-scale dynamic feature fusion approach for modern network architectures. Our work in a nutshell is like a Russian doll. We propose a nested optimization on the fusion of received features. Meanwhile, we advocate the idea that channel attention can also have a choice of scale, and a multi-scale channel attention mechanism is introduced to handle the issue of object scale variation.

Yimian Dai

Postdoctoral Fellow

My research interests include image restoration, object detection, and vision-language models.