当前位置：网站首页>Bisenetv2 face segmentation

Bisenetv2 face segmentation

2022-07-19 03:13:00 【HySmiley】

One 、 The paper

BiSeNet V2: Bilateral Network with Guided Aggregation for Real-time Semantic Segmentation

https://arxiv.org/abs/2004.02147

BiSeNet V2 Separate these spatial details from the classification semantics , To achieve high-precision and efficient real-time semantic segmentation . So , An effective architecture is proposed , Trade off between speed and accuracy , It's called a bilateral segmentation network (BiSeNet V2). The architecture includes :(1) A branch of detail , With wide channels and shallow layers , Used to capture low-level details and generate high-resolution feature representations ;(2) A semantic Branch , Narrow passage , Deep level , Obtaining high level semantic context . Semantic branching is lightweight , Because it reduces the channel capacity and fast down sampling strategy . Besides , A guided aggregation layer is designed to enhance the interconnection and fusion of these two types of feature representation . Besides , An enhanced training strategy is also designed , It can improve the segmentation performance without increasing any inference cost .