【语义分割】Fast-SCNN -- Fast Semantic Segmentation Network

paper：Fast-SCNN
github：code

---

Abstract

Fast-SCNN是一个实时的语义分割模型。其基于现有的two-branch方法（BiSeNet），引入了一个learning to downsample模块，在cityscapes上得到68.0%的miou。

FastSCNN采用depthwise separable convolutions和inverse residual blocks

Fast-SCNN

实时语义分割模型设计要点：

• a larger receptive field is important to learn complex correlations among object classes (i.e. global context)
• spatial detail in images is necessary to preserve object
  boundaries
• balance speed and accuracy

Network Architecture

PPM是PSPNet中的一个重要模块，可融合 different-region-based context information.

Adding few layers after the feature fusion module boosts the accuracy

Pre-training on Auxiliary Tasks

In our experiments we show that small networks do not get significant benefit from pre-training. Instead, aggressive data augmentation and more number of epochs provide similar results

Fast-SCNN在cityscapes上训练1000个epoch

Experiment

1、Evaluation on Cityscapes

2、Pre-training and Weakly Labeled Data

3、Lower Input Resolution

目前很多模型输入都为512 x 1024